Memory-optimized shuffle file consolidation
Reduces overhead of each shuffle block for consolidation from >300 bytes to 8 bytes (1 primitive Long). Verified via profiler testing with 1 mil shuffle blocks, net overhead was ~8,400,000 bytes.
Despite the memory-optimized implementation incurring extra CPU overhead, the runtime of the shuffle phase in this test was only around 2% slower, while the reduce phase was 40% faster, when compared to not using any shuffle file consolidation.
This is accomplished by replacing the map from ShuffleBlockId to FileSegment (i.e., block id to where it's located), which had high overhead due to being a gigantic, timestamped, concurrent map with a more space-efficient structure. Namely, the following are introduced (I have omitted the word "Shuffle" from some names for clarity):
**ShuffleFile** - there is one ShuffleFile per consolidated shuffle file on disk. We store an array of offsets into the physical shuffle file for each ShuffleMapTask that wrote into the file. This is sufficient to reconstruct FileSegments for mappers that are in the file.
**FileGroup** - contains a set of ShuffleFiles, one per reducer, that a MapTask can use to write its output. There is one FileGroup created per _concurrent_ MapTask. The FileGroup contains an array of the mapIds that have been written to all files in the group. The positions of elements in this array map directly onto the positions in each ShuffleFile's offsets array.
In order to locate the FileSegment associated with a BlockId, we have another structure which maps each reducer to the set of ShuffleFiles that were created for it. (There will be as many ShuffleFiles per reducer as there are FileGroups.) To lookup a given ShuffleBlockId (shuffleId, reducerId, mapId), we thus search through all ShuffleFiles associated with that reducer.
As a time optimization, we ensure that FileGroups are only reused for MapTasks with monotonically increasing mapIds. This allows us to perform a binary search to locate a mapId inside a group, and also enables potential future optimization (based on the usual monotonic access order).
- ShuffleBlocks has been removed and replaced by ShuffleWriterGroup.
- ShuffleWriterGroup no longer contains a reference to a ShuffleFileGroup.
- ShuffleFile has been removed and its contents are now within ShuffleFileGroup.
- ShuffleBlockManager.forShuffle has been replaced by a more stateful forMapTask.
For some reason, even calling
java.nio.Files.createTempDirectory().getFile.deleteOnExit()
does not delete the directory on exit. Guava's analagous function
seems to work, however.
Overhead of each shuffle block for consolidation has been reduced from >300 bytes
to 8 bytes (1 primitive Long). Verified via profiler testing with 1 mil shuffle blocks,
net overhead was ~8,400,000 bytes.
Despite the memory-optimized implementation incurring extra CPU overhead, the runtime
of the shuffle phase in this test was only around 2% slower, while the reduce phase
was 40% faster, when compared to not using any shuffle file consolidation.
Fast, memory-efficient hash set, hash table implementations optimized for primitive data types.
This pull request adds two hash table implementations optimized for primitive data types. For primitive types, the new hash tables are much faster than the current Spark AppendOnlyMap (3X faster - note that the current AppendOnlyMap is already much better than the Java map) while uses much less space (1/4 of the space).
Details:
This PR first adds a open hash set implementation (OpenHashSet) optimized for primitive types (using Scala's specialization feature). This OpenHashSet is designed to serve as building blocks for more advanced structures. It is currently used to build the following two hash tables, but can be used in the future to build multi-valued hash tables as well (GraphX has this use case). Note that there are some peculiarities in the code for working around some Scala compiler bugs.
Building on top of OpenHashSet, this PR adds two different hash tables implementations:
1. OpenHashSet: for nullable keys, optional specialization for primitive values
2. PrimitiveKeyOpenHashMap: for primitive keys that are not nullable, and optional specialization for primitive values
I tested the update speed of these two implementations using the changeValue function (which is what Aggregator and cogroup would use). Runtime relative to AppendOnlyMap for inserting 10 million items:
Int to Int: ~30%
java.lang.Integer to java.lang.Integer: ~100%
Int to java.lang.Integer: ~50%
java.lang.Integer to Int: ~85%
Document & finish support for local: URIs
Review all the supported URI schemes for addJar / addFile to the Cluster Overview page.
Add support for local: URI to addFile.
Handle ConcurrentModificationExceptions in SparkContext init.
System.getProperties.toMap will fail-fast when concurrently modified,
and it seems like some other thread started by SparkContext does
a System.setProperty during it's initialization.
Handle this by just looping on ConcurrentModificationException, which
seems the safest, since the non-fail-fast methods (Hastable.entrySet)
have undefined behavior under concurrent modification.
The code in LocalScheduler/LocalTaskSetManager was nearly identical
to the code in ClusterScheduler/ClusterTaskSetManager. The redundancy
made making updating the schedulers unnecessarily painful and error-
prone. This commit combines the two into a single TaskScheduler/
TaskSetManager.
Fixed incorrect log message in local scheduler
This change is especially relevant at the moment, because some users are seeing this failure, and the log message is misleading/incorrect (because for the tests, the max failures is set to 0, not 4)
Pull SparkHadoopUtil out of SparkEnv (jira SPARK-886)
Having the logic to initialize the correct SparkHadoopUtil in SparkEnv prevents it from being used until after the SparkContext is initialized. This causes issues like https://spark-project.atlassian.net/browse/SPARK-886. It also makes it hard to use in singleton objects. For instance I want to use it in the security code.
Add support for local:// URI scheme for addJars()
This PR adds support for a new URI scheme for SparkContext.addJars(): `local://file/path`.
The *local* scheme indicates that the `/file/path` exists on every worker node. The reason for its existence is for big library JARs, which would be really expensive to serve using the standard HTTP fileserver distribution method, especially for big clusters. Today the only inexpensive method (assuming such a file is on every host, via say NFS, rsync, etc.) of doing this is to add the JAR to the SPARK_CLASSPATH, but we want a method where the user does not need to modify the Spark configuration.
I would add something to the docs, but it's not obvious where to add it.
Oh, and it would be great if this could be merged in time for 0.8.1.
Reduce the memory footprint of BlockInfo objects
This pull request reduces the memory footprint of all BlockInfo objects and makes additional optimizations for shuffle blocks. For all BlockInfo objects, these changes remove two boolean fields and one Object field. For shuffle blocks, we additionally remove an Object field and a boolean field.
When storing tens of thousands of these objects, this may add up to significant memory savings. A ShuffleBlockInfo now only needs to wrap a single long.
This was motivated by a [report of high blockInfo memory usage during shuffles](https://mail-archives.apache.org/mod_mbox/incubator-spark-user/201310.mbox/%3C20131026134353.202b2b9b%40sh9%3E).
I haven't run benchmarks to measure the exact memory savings.
/cc @aarondav
Display both task ID and task attempt ID in UI, and rename taskId to taskAttemptId
Previously only the task attempt ID was shown in the UI; this was confusing because the job can be shown as complete while there are tasks still running. Showing the task ID in addition to the attempt ID makes it clear which tasks are redundant.
This commit also renames taskId to taskAttemptId in TaskInfo and in the local/cluster schedulers. This identifier was used to uniquely identify attempts, not tasks, so the current naming was confusing. The new naming is also more consistent with map reduce.
Eliminate extra memory usage when shuffle file consolidation is disabled
Otherwise, we see SPARK-946 even when shuffle file consolidation is disabled.
Fixing SPARK-946 is still forthcoming.
System.getProperties.toMap will fail-fast when concurrently modified,
and it seems like some other thread started by SparkContext does
a System.setProperty during it's initialization.
Handle this by just looping on ConcurrentModificationException, which
seems the safest, since the non-fail-fast methods (Hastable.entrySet)
have undefined behavior under concurrent modification.