ODIn/spark-instrumented-optimizer
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity
spark-instrumented-optimizer/sql/core/benchmarks/InExpressionBenchmark-results.txt

741 lines
64 KiB
Plaintext
Raw Normal View History

[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
================================================================================================
In Expression Benchmark
================================================================================================
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
5 bytes: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 132 172 32 75.7 13.2 1.0X
InSet expression 79 98 13 125.8 7.9 1.7X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
10 bytes: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 100 111 7 99.7 10.0 1.0X
InSet expression 70 78 9 143.0 7.0 1.4X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
25 bytes: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 161 170 7 62.0 16.1 1.0X
InSet expression 88 93 7 113.9 8.8 1.8X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
50 bytes: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 270 277 6 37.0 27.0 1.0X
InSet expression 116 123 9 86.0 11.6 2.3X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
100 bytes: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 450 463 13 22.2 45.0 1.0X
InSet expression 182 189 7 54.9 18.2 2.5X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
200 bytes: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 706 714 7 14.2 70.6 1.0X
InSet expression 302 311 7 33.1 30.2 2.3X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
5 shorts: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 62 64 5 162.2 6.2 1.0X
InSet expression 57 59 5 176.7 5.7 1.1X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
10 shorts: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 79 82 5 126.7 7.9 1.0X
InSet expression 53 54 3 188.7 5.3 1.5X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
25 shorts: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 132 135 5 75.8 13.2 1.0X
InSet expression 50 53 6 200.7 5.0 2.6X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
50 shorts: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 228 231 4 43.8 22.8 1.0X
InSet expression 51 53 3 195.5 5.1 4.5X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
100 shorts: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 404 404 0 24.8 40.4 1.0X
InSet expression 58 61 4 171.0 5.8 6.9X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
200 shorts: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 766 767 1 13.1 76.6 1.0X
InSet expression 66 68 3 151.0 6.6 11.6X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
300 shorts: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1124 1124 0 8.9 112.4 1.0X
InSet expression 74 77 4 135.3 7.4 15.2X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
400 shorts: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1566 1567 2 6.4 156.6 1.0X
InSet expression 82 84 4 121.9 8.2 19.1X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
500 shorts: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1841 1871 37 5.4 184.1 1.0X
InSet expression 314 318 5 31.8 31.4 5.9X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
5 shorts (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 58 62 5 171.0 5.8 1.0X
InSet expression 53 55 4 187.5 5.3 1.1X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
10 shorts (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 76 77 2 131.3 7.6 1.0X
InSet expression 65 66 3 154.3 6.5 1.2X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
25 shorts (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 138 140 5 72.7 13.8 1.0X
InSet expression 74 78 8 135.2 7.4 1.9X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
50 shorts (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 226 227 1 44.2 22.6 1.0X
InSet expression 83 86 7 120.8 8.3 2.7X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
100 shorts (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 402 403 0 24.9 40.2 1.0X
InSet expression 93 94 3 108.0 9.3 4.3X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
200 shorts (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 761 762 0 13.1 76.1 1.0X
InSet expression 113 116 7 88.4 11.3 6.7X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
300 shorts (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1125 1125 0 8.9 112.5 1.0X
InSet expression 136 142 11 73.5 13.6 8.3X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
400 shorts (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1486 1487 1 6.7 148.6 1.0X
InSet expression 141 142 2 70.8 14.1 10.5X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
500 shorts (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1842 1873 67 5.4 184.2 1.0X
InSet expression 315 318 3 31.7 31.5 5.8X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
5 ints: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 54 55 4 186.6 5.4 1.0X
InSet expression 49 51 3 203.0 4.9 1.1X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
10 ints: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 72 74 5 139.7 7.2 1.0X
InSet expression 46 48 5 218.2 4.6 1.6X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
25 ints: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 125 127 5 79.9 12.5 1.0X
InSet expression 47 48 4 212.5 4.7 2.7X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
50 ints: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 221 224 4 45.2 22.1 1.0X
InSet expression 48 49 3 206.3 4.8 4.6X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
100 ints: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 401 404 6 25.0 40.1 1.0X
InSet expression 55 56 2 180.5 5.5 7.2X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
200 ints: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 762 763 1 13.1 76.2 1.0X
InSet expression 63 69 14 159.8 6.3 12.2X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
300 ints: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1117 1117 0 9.0 111.7 1.0X
InSet expression 70 71 2 143.3 7.0 16.0X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
400 ints: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1557 1558 1 6.4 155.7 1.0X
InSet expression 77 78 2 129.6 7.7 20.2X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
500 ints: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1841 1877 79 5.4 184.1 1.0X
InSet expression 320 322 2 31.2 32.0 5.8X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
5 ints (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 43 44 3 231.6 4.3 1.0X
InSet expression 40 42 4 252.4 4.0 1.1X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
10 ints (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 62 64 4 162.0 6.2 1.0X
InSet expression 45 47 4 222.2 4.5 1.4X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
25 ints (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 117 119 4 85.4 11.7 1.0X
InSet expression 57 59 5 176.9 5.7 2.1X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
50 ints (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 207 207 0 48.3 20.7 1.0X
InSet expression 65 66 3 153.3 6.5 3.2X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
100 ints (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 390 394 4 25.7 39.0 1.0X
InSet expression 76 77 3 132.0 7.6 5.1X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
200 ints (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 752 755 2 13.3 75.2 1.0X
InSet expression 111 112 3 90.2 11.1 6.8X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
300 ints (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1106 1115 6 9.0 110.6 1.0X
InSet expression 129 130 3 77.7 12.9 8.6X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
400 ints (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1476 1484 6 6.8 147.6 1.0X
InSet expression 129 130 3 77.3 12.9 11.4X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
500 ints (non-compact): Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1844 1900 124 5.4 184.4 1.0X
InSet expression 321 322 2 31.2 32.1 5.7X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
5 longs: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 50 53 5 199.5 5.0 1.0X
InSet expression 166 169 5 60.2 16.6 0.3X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
10 longs: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 64 68 8 155.5 6.4 1.0X
InSet expression 186 188 4 53.9 18.6 0.3X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
25 longs: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 118 119 2 84.7 11.8 1.0X
InSet expression 194 208 26 51.4 19.4 0.6X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
50 longs: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 208 208 0 48.1 20.8 1.0X
InSet expression 240 244 5 41.7 24.0 0.9X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
100 longs: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 390 397 10 25.6 39.0 1.0X
InSet expression 205 207 5 48.8 20.5 1.9X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
200 longs: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 745 745 1 13.4 74.5 1.0X
InSet expression 194 197 5 51.5 19.4 3.8X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
5 floats: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 90 91 1 111.2 9.0 1.0X
InSet expression 199 202 4 50.2 19.9 0.5X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
10 floats: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 131 132 0 76.4 13.1 1.0X
InSet expression 221 223 2 45.2 22.1 0.6X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
25 floats: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 244 245 0 40.9 24.4 1.0X
InSet expression 235 236 1 42.6 23.5 1.0X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
50 floats: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 423 424 2 23.7 42.3 1.0X
InSet expression 284 285 2 35.3 28.4 1.5X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
100 floats: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 777 778 1 12.9 77.7 1.0X
InSet expression 248 249 2 40.4 24.8 3.1X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
200 floats: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 3032 3125 203 3.3 303.2 1.0X
InSet expression 239 241 2 41.8 23.9 12.7X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
5 doubles: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 84 86 3 118.8 8.4 1.0X
InSet expression 167 168 2 59.9 16.7 0.5X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
10 doubles: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 126 127 1 79.1 12.6 1.0X
InSet expression 183 185 2 54.6 18.3 0.7X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
25 doubles: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 239 240 0 41.8 23.9 1.0X
InSet expression 189 192 4 52.9 18.9 1.3X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
50 doubles: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 417 417 0 24.0 41.7 1.0X
InSet expression 231 234 4 43.3 23.1 1.8X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
100 doubles: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 770 772 3 13.0 77.0 1.0X
InSet expression 201 204 5 49.7 20.1 3.8X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
200 doubles: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 3587 3686 212 2.8 358.7 1.0X
InSet expression 196 198 3 50.9 19.6 18.3X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
5 small decimals: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 50 51 2 20.1 49.7 1.0X
InSet expression 151 153 3 6.6 150.7 0.3X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
10 small decimals: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 62 62 2 16.2 61.6 1.0X
InSet expression 153 155 2 6.5 153.3 0.4X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
25 small decimals: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 95 96 2 10.5 95.0 1.0X
InSet expression 156 158 2 6.4 156.4 0.6X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
50 small decimals: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 175 177 5 5.7 175.2 1.0X
InSet expression 165 167 3 6.1 164.9 1.1X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
100 small decimals: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 440 440 0 2.3 439.7 1.0X
InSet expression 167 169 2 6.0 167.0 2.6X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
200 small decimals: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1005 1154 330 1.0 1004.8 1.0X
InSet expression 180 182 2 5.5 180.2 5.6X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
5 large decimals: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 251 258 7 4.0 251.1 1.0X
InSet expression 193 197 3 5.2 193.2 1.3X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
10 large decimals: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 424 437 8 2.4 424.2 1.0X
InSet expression 196 198 2 5.1 195.5 2.2X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
25 large decimals: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1018 1023 5 1.0 1017.8 1.0X
InSet expression 203 206 3 4.9 202.9 5.0X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
50 large decimals: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1947 1955 7 0.5 1947.5 1.0X
InSet expression 208 211 3 4.8 208.4 9.3X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
100 large decimals: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 3886 3899 10 0.3 3885.9 1.0X
InSet expression 233 235 4 4.3 232.6 16.7X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
200 large decimals: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 7702 7793 60 0.1 7701.8 1.0X
InSet expression 243 248 6 4.1 243.4 31.6X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
5 strings: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 132 135 4 7.6 132.1 1.0X
InSet expression 147 149 2 6.8 147.1 0.9X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
10 strings: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 139 141 5 7.2 139.0 1.0X
InSet expression 150 151 2 6.7 149.6 0.9X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
25 strings: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 160 161 2 6.3 159.6 1.0X
InSet expression 157 158 2 6.4 157.3 1.0X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
50 strings: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 190 192 2 5.3 189.8 1.0X
InSet expression 160 161 2 6.2 160.5 1.2X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
100 strings: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 264 265 2 3.8 263.8 1.0X
InSet expression 159 160 2 6.3 158.5 1.7X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
200 strings: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 735 884 327 1.4 735.4 1.0X
InSet expression 164 166 3 6.1 163.9 4.5X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
5 timestamps: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 45 46 3 223.6 4.5 1.0X
InSet expression 162 164 3 61.6 16.2 0.3X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
10 timestamps: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 62 65 7 161.9 6.2 1.0X
InSet expression 179 181 2 56.0 17.9 0.3X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
25 timestamps: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 115 116 1 86.8 11.5 1.0X
InSet expression 222 225 6 45.1 22.2 0.5X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
50 timestamps: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 198 199 1 50.6 19.8 1.0X
InSet expression 238 239 2 42.0 23.8 0.8X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
100 timestamps: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 363 368 5 27.6 36.3 1.0X
InSet expression 222 224 2 45.0 22.2 1.6X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
200 timestamps: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 701 706 4 14.3 70.1 1.0X
InSet expression 226 228 3 44.2 22.6 3.1X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
5 dates: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 659 665 5 15.2 65.9 1.0X
InSet expression 660 664 3 15.1 66.0 1.0X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
10 dates: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 663 667 2 15.1 66.3 1.0X
InSet expression 659 661 2 15.2 65.9 1.0X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
25 dates: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 700 705 4 14.3 70.0 1.0X
InSet expression 667 669 2 15.0 66.7 1.0X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
50 dates: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 801 805 3 12.5 80.1 1.0X
InSet expression 675 677 1 14.8 67.5 1.2X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
100 dates: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 984 987 5 10.2 98.4 1.0X
InSet expression 685 690 4 14.6 68.5 1.4X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
200 dates: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1350 1356 4 7.4 135.0 1.0X
InSet expression 710 712 3 14.1 71.0 1.9X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
300 dates: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1716 1722 6 5.8 171.6 1.0X
InSet expression 716 719 2 14.0 71.6 2.4X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
400 dates: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 2083 2100 13 4.8 208.3 1.0X
InSet expression 742 744 1 13.5 74.2 2.8X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
500 dates: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 2469 2486 11 4.1 246.9 1.0X
InSet expression 829 831 2 12.1 82.9 3.0X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
5 arrays: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 56 58 4 17.9 55.9 1.0X
InSet expression 123 124 2 8.1 123.1 0.5X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
10 arrays: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 83 84 2 12.1 82.7 1.0X
InSet expression 124 127 7 8.1 123.8 0.7X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
25 arrays: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 275 281 9 3.6 275.5 1.0X
InSet expression 155 158 4 6.4 155.2 1.8X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
50 arrays: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 659 718 109 1.5 659.4 1.0X
InSet expression 217 218 2 4.6 217.0 3.0X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
100 arrays: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 2488 2684 434 0.4 2488.4 1.0X
InSet expression 267 270 3 3.7 266.7 9.3X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
200 arrays: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 9462 10091 897 0.1 9462.2 1.0X
InSet expression 347 349 2 2.9 347.4 27.2X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
5 structs: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 51 55 7 19.7 50.8 1.0X
InSet expression 166 168 4 6.0 166.3 0.3X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
10 structs: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 66 70 5 15.1 66.4 1.0X
InSet expression 167 170 3 6.0 167.3 0.4X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
25 structs: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 142 144 3 7.0 142.1 1.0X
InSet expression 211 215 6 4.7 211.2 0.7X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
50 structs: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 375 378 5 2.7 375.5 1.0X
InSet expression 297 298 2 3.4 297.5 1.3X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
100 structs: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 1122 1260 304 0.9 1122.2 1.0X
InSet expression 354 357 4 2.8 353.9 3.2X
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
OpenJDK 64-Bit Server VM 1.8.0_232-8u232-b09-0ubuntu1~18.04.1-b09 on Linux 4.15.0-1044-aws
[SPARK-26205][SQL] Optimize InSet Expression for bytes, shorts, ints, dates ## What changes were proposed in this pull request? This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai. `In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons. The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\). Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements. I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`). See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information. This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search. ## How was this patch tested? There are new tests that verify the logic of the proposed optimization. The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz). ## Notes - [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method. Closes #23171 from aokolnychyi/spark-26205. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2019-03-04 18:40:04 -05:00
Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
200 structs: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
[SPARK-30409][SPARK-29173][SQL][TESTS] Use `NoOp` datasource in SQL benchmarks ### What changes were proposed in this pull request? In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`. ### Why are the changes needed? To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations. ### Does this PR introduce any user-facing change? No ### How was this patch tested? Re-run all modified benchmarks using Amazon EC2. | Item | Description | | ---- | ----| | Region | us-west-2 (Oregon) | | Instance | r3.xlarge (spot instance) | | AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) | | Java | OpenJDK8/10 | - Run `TPCDSQueryBenchmark` using instructions from the PR #26049 ``` # `spark-tpcds-datagen` needs this. (JDK8) $ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4 $ export SPARK_HOME=$PWD $ ./build/mvn clean package -DskipTests # Generate data. (JDK8) $ git clone gitgithub.com:maropu/spark-tpcds-datagen.git $ cd spark-tpcds-datagen/ $ build/mvn clean package $ mkdir -p /data/tpcds $ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4` ``` - Other benchmarks ran by the script: ``` #!/usr/bin/env python3 import os from sparktestsupport.shellutils import run_cmd benchmarks = [ ['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'], ['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'], ['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'], ['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'], ['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark'] ] print('Set SPARK_GENERATE_BENCHMARK_FILES=1') os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1' for b in benchmarks: print("Run benchmark: %s" % b[1]) run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])]) ``` Closes #27078 from MaxGekk/noop-in-benchmarks. Lead-authored-by: Maxim Gekk <max.gekk@gmail.com> Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com> Co-authored-by: Dongjoon Hyun <dhyun@apple.com> Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
2020-01-12 16:18:19 -05:00
In expression 5651 5989 644 0.2 5651.4 1.0X
InSet expression 471 473 2 2.1 471.0 12.0X
[SPARK-26203][SQL][TEST] Benchmark performance of In and InSet expressions ## What changes were proposed in this pull request? This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints. As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?). ### byte / short / int / long Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing . Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode. ### float / double Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements. Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`). ### decimal The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases. If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive. `Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`. ### string `UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases. ### timestamp / date Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing. ### array Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`. ### struct `InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`. Closes #23291 from aokolnychyi/spark-26203. Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com> Co-authored-by: Dongjoon Hyun <dongjoon@apache.org> Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
2019-01-15 09:25:50 -05:00
Reference in a new issue Copy permalink