### 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>
## 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>
## 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>
Maxim Gekk
Anton Okolnychyi