New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[SPARK-7076][SPARK-7077][SPARK-7080][SQL] Use managed memory for aggregations #5725

Closed
wants to merge 59 commits into
base: master
from

Conversation

Projects
None yet
9 participants
@JoshRosen
Contributor

JoshRosen commented Apr 27, 2015

This patch adds managed-memory-based aggregation to Spark SQL / DataFrames. Instead of working with Java objects, this new aggregation path uses sun.misc.Unsafe to manipulate raw memory. This reduces the memory footprint for aggregations, resulting in fewer spills, OutOfMemoryErrors, and garbage collection pauses. As a result, this allows for higher memory utilization. It can also result in better cache locality since objects will be stored closer together in memory.

This feature can be eanbled by setting spark.sql.unsafe.enabled=true. For now, this feature is only supported when codegen is enabled and only supports aggregations for which the grouping columns are primitive numeric types or strings and aggregated values are numeric.

Managing memory with sun.misc.Unsafe

This patch supports both on- and off-heap managed memory.

  • In on-heap mode, memory addresses are identified by the combination of a base Object and an offset within that object.
  • In off-heap mode, memory is addressed directly with 64-bit long addresses.

To support both modes, functions that manipulate memory accept both baseObject and baseOffset fields. In off-heap mode, we simply pass null as baseObject.

We allocate memory in large chunks, so memory fragmentation and allocation speed are not significant bottlenecks.

By default, we use on-heap mode. To enable off-heap mode, set spark.unsafe.offHeap=true.

To track allocated memory, this patch extends SparkEnv with an ExecutorMemoryManager and supplies each TaskContext with a TaskMemoryManager. These classes work together to track allocations and detect memory leaks.

Compact tuple format

This patch introduces UnsafeRow, a compact row layout. In this format, each tuple has three parts: a null bit set, fixed length values, and variable-length values:

image

  • Rows are always 8-byte word aligned (so their sizes will always be a multiple of 8 bytes)
  • The bit set is used for null tracking:
    • Position i is set if and only if field i is null
    • The bit set is aligned to an 8-byte word boundary.
  • Every field appears as an 8-byte word in the fixed-length values part:
    • If a field is null, we zero out the values.
    • If a field is variable-length, the word stores a relative offset (w.r.t. the base of the tuple) that points to the beginning of the field's data in the variable-length part.
  • Each variable-length data type can have its own encoding:
    • For strings, the first word stores the length of the string and is followed by UTF-8 encoded bytes. If necessary, the end of the string is padded with empty bytes in order to ensure word-alignment.

For example, a tuple that consists 3 fields of type (int, string, string), with value (null, “data”, “bricks”) would look like this:

image

This format allows us to compare tuples for equality by directly comparing their raw bytes. This also enables fast hashing of tuples.

Hash map for performing aggregations

This patch introduces UnsafeFixedWidthAggregationMap, a hash map for performing aggregations where the aggregation result columns are fixed-with. This map's keys and values are Row objects. UnsafeFixedWidthAggregationMap is implemented on top of BytesToBytesMap, an append-only map which supports byte-array keys and values.

BytesToBytesMap stores pointers to key and value tuples. For each record with a new key, we copy the key and create the aggregation value buffer for that key and put them in a buffer. The hash table then simply stores pointers to the key and value. For each record with an existing key, we simply run the aggregation function to update the values in place.

This map is implemented using open hashing with triangular sequence probing. Each entry stores two words in a long array: the first word stores the address of the key and the second word stores the relative offset from the key tuple to the value tuple, as well as the key's 32-bit hashcode. By storing the full hashcode, we reduce the number of equality checks that need to be performed to handle position collisions ()since the chance of hashcode collision is much lower than position collision).

UnsafeFixedWidthAggregationMap allows regular Spark SQL Row objects to be used when probing the map. Internally, it encodes these rows into UnsafeRow format using UnsafeRowConverter. This conversion has a small overhead that can be eliminated in the future once we use UnsafeRows in other operators.

Review on Reviewable

JoshRosen added some commits Apr 17, 2015

Add skeleton for GeneratedAggregate integration.
This typechecks properly and sketches how I'm intending to use row pointers
and the hashmap. This has been a useful exercise for figuring out whether my
interfaces will be sufficient.
First version that passes some aggregation tests:
I commented out a number of tests where we do not support the required
data types; this is only a short-term hack until I extend the planner
to understand when UnsafeGeneratedAggregate can be used.
Fix bug in UnsafeMemoryAllocator.free():
The `if` check here was backwards, which prevented any memory from
being freed.
Optimizations related to zeroing out memory:
- Do not zero out all allocated memory; the zeroing isn't free and in many
  cases it isn't necessary.
- There are some cases where we do want to clear the memory, such as in BitSet.

  It shouldn't be the BitSet object's responsibility to zero out the memory
  block passed to it (since maybe we're passing some memory created by someone
  else and want to interpret it as a bitset). To make the caller's life easier,
  though, I added a MemoryBlock.zero() method for clearing the block.

- In UnsafeGeneratedAggregate, use Arrays.fill to clear the re-used temporary
  row buffer, since this is likely to be much faster than Unsafe.setMemory;
  see http://psy-lob-saw.blogspot.com/2015/04/on-arraysfill-intrinsics-superword-and.html
  for more details.
Extract aggregation map into its own class.
This makes the code much easier to understand and
will allow me to implement unsafe versions of both
GeneratedAggregate and the regular Aggregate operator.
Merge Unsafe code into the regular GeneratedAggregate, guarded by a
configuration flag; integrate planner support and re-enable all tests.
Cleanups to string handling code
Changes to UTF8String will be deferred for a followup, since we should
benchmark them first.
@SparkQA

This comment has been minimized.

Show comment
Hide comment
@SparkQA

SparkQA commented Apr 29, 2015

Test build #31202 has started for PR 5725 at commit eeee512.

@rxin

This comment has been minimized.

Show comment
Hide comment
@rxin

rxin Apr 29, 2015

Contributor

LGTM. There are some issues, but we can easily address them in followup PRs.

Contributor

rxin commented Apr 29, 2015

LGTM. There are some issues, but we can easily address them in followup PRs.

@SparkQA

This comment has been minimized.

Show comment
Hide comment
@SparkQA

SparkQA Apr 29, 2015

Test build #31202 has finished for PR 5725 at commit eeee512.

  • This patch fails MiMa tests.
  • This patch merges cleanly.
  • This patch adds the following public classes (experimental):
    • public final class UnsafeFixedWidthAggregationMap
    • public static class MapEntry
    • public final class UnsafeRow implements MutableRow
    • class UnsafeRowConverter(fieldTypes: Array[DataType])
    • public final class PlatformDependent
    • public class ByteArrayMethods
    • public final class LongArray
    • public final class BitSet
    • public final class BitSetMethods
    • public final class Murmur3_x86_32
    • public final class BytesToBytesMap
    • public final class Location
    • class Doubling implements HashMapGrowthStrategy
    • public class ExecutorMemoryManager
    • public class HeapMemoryAllocator implements MemoryAllocator
    • public class MemoryBlock extends MemoryLocation
    • public class MemoryLocation
    • public final class TaskMemoryManager
    • public class UnsafeMemoryAllocator implements MemoryAllocator
  • This patch adds the following new dependencies:
    • spark-unsafe_2.10-1.4.0-SNAPSHOT.jar

SparkQA commented Apr 29, 2015

Test build #31202 has finished for PR 5725 at commit eeee512.

  • This patch fails MiMa tests.
  • This patch merges cleanly.
  • This patch adds the following public classes (experimental):
    • public final class UnsafeFixedWidthAggregationMap
    • public static class MapEntry
    • public final class UnsafeRow implements MutableRow
    • class UnsafeRowConverter(fieldTypes: Array[DataType])
    • public final class PlatformDependent
    • public class ByteArrayMethods
    • public final class LongArray
    • public final class BitSet
    • public final class BitSetMethods
    • public final class Murmur3_x86_32
    • public final class BytesToBytesMap
    • public final class Location
    • class Doubling implements HashMapGrowthStrategy
    • public class ExecutorMemoryManager
    • public class HeapMemoryAllocator implements MemoryAllocator
    • public class MemoryBlock extends MemoryLocation
    • public class MemoryLocation
    • public final class TaskMemoryManager
    • public class UnsafeMemoryAllocator implements MemoryAllocator
  • This patch adds the following new dependencies:
    • spark-unsafe_2.10-1.4.0-SNAPSHOT.jar
@AmplabJenkins

This comment has been minimized.

Show comment
Hide comment
@AmplabJenkins

AmplabJenkins Apr 29, 2015

Merged build finished. Test FAILed.

AmplabJenkins commented Apr 29, 2015

Merged build finished. Test FAILed.

@AmplabJenkins

This comment has been minimized.

Show comment
Hide comment
@AmplabJenkins

AmplabJenkins Apr 29, 2015

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/31202/
Test FAILed.

AmplabJenkins commented Apr 29, 2015

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/31202/
Test FAILed.

@JoshRosen

This comment has been minimized.

Show comment
Hide comment
@JoshRosen

JoshRosen Apr 29, 2015

Contributor

Jenkins, retest this please.

Contributor

JoshRosen commented Apr 29, 2015

Jenkins, retest this please.

@AmplabJenkins

This comment has been minimized.

Show comment
Hide comment
@AmplabJenkins

AmplabJenkins Apr 29, 2015

Merged build triggered.

AmplabJenkins commented Apr 29, 2015

Merged build triggered.

@AmplabJenkins

This comment has been minimized.

Show comment
Hide comment
@AmplabJenkins

AmplabJenkins Apr 29, 2015

Merged build started.

AmplabJenkins commented Apr 29, 2015

Merged build started.

@SparkQA

This comment has been minimized.

Show comment
Hide comment
@SparkQA

SparkQA commented Apr 29, 2015

Test build #31245 has started for PR 5725 at commit eeee512.

@SparkQA

This comment has been minimized.

Show comment
Hide comment
@SparkQA

SparkQA commented Apr 29, 2015

Test build #732 has started for PR 5725 at commit eeee512.

@SparkQA

This comment has been minimized.

Show comment
Hide comment
@SparkQA

SparkQA Apr 29, 2015

Test build #31245 has finished for PR 5725 at commit eeee512.

  • This patch passes all tests.
  • This patch merges cleanly.
  • This patch adds the following public classes (experimental):
    • public final class UnsafeFixedWidthAggregationMap
    • public static class MapEntry
    • public final class UnsafeRow implements MutableRow
    • class UnsafeRowConverter(fieldTypes: Array[DataType])
    • public final class PlatformDependent
    • public class ByteArrayMethods
    • public final class LongArray
    • public final class BitSet
    • public final class BitSetMethods
    • public final class Murmur3_x86_32
    • public final class BytesToBytesMap
    • public final class Location
    • class Doubling implements HashMapGrowthStrategy
    • public class ExecutorMemoryManager
    • public class HeapMemoryAllocator implements MemoryAllocator
    • public class MemoryBlock extends MemoryLocation
    • public class MemoryLocation
    • public final class TaskMemoryManager
    • public class UnsafeMemoryAllocator implements MemoryAllocator
  • This patch adds the following new dependencies:
    • spark-unsafe_2.10-1.4.0-SNAPSHOT.jar

SparkQA commented Apr 29, 2015

Test build #31245 has finished for PR 5725 at commit eeee512.

  • This patch passes all tests.
  • This patch merges cleanly.
  • This patch adds the following public classes (experimental):
    • public final class UnsafeFixedWidthAggregationMap
    • public static class MapEntry
    • public final class UnsafeRow implements MutableRow
    • class UnsafeRowConverter(fieldTypes: Array[DataType])
    • public final class PlatformDependent
    • public class ByteArrayMethods
    • public final class LongArray
    • public final class BitSet
    • public final class BitSetMethods
    • public final class Murmur3_x86_32
    • public final class BytesToBytesMap
    • public final class Location
    • class Doubling implements HashMapGrowthStrategy
    • public class ExecutorMemoryManager
    • public class HeapMemoryAllocator implements MemoryAllocator
    • public class MemoryBlock extends MemoryLocation
    • public class MemoryLocation
    • public final class TaskMemoryManager
    • public class UnsafeMemoryAllocator implements MemoryAllocator
  • This patch adds the following new dependencies:
    • spark-unsafe_2.10-1.4.0-SNAPSHOT.jar
@AmplabJenkins

This comment has been minimized.

Show comment
Hide comment
@AmplabJenkins

AmplabJenkins Apr 29, 2015

Merged build finished. Test PASSed.

AmplabJenkins commented Apr 29, 2015

Merged build finished. Test PASSed.

@AmplabJenkins

This comment has been minimized.

Show comment
Hide comment
@AmplabJenkins

AmplabJenkins Apr 29, 2015

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/31245/
Test PASSed.

AmplabJenkins commented Apr 29, 2015

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/31245/
Test PASSed.

@rxin

This comment has been minimized.

Show comment
Hide comment
@rxin

rxin Apr 29, 2015

Contributor

Ok merging in master. Let's make sure we address @marmbrus 's comments in followup prs.

Contributor

rxin commented Apr 29, 2015

Ok merging in master. Let's make sure we address @marmbrus 's comments in followup prs.

@asfgit asfgit closed this in f49284b Apr 29, 2015

@tedyu
@tgravescs

This comment has been minimized.

Show comment
Hide comment
@tgravescs

tgravescs Apr 30, 2015

Contributor

I'm not able to build after this went in. Any special requirements to build this?

[info] Compiling 15 Java sources to /home/tgraves/tgravescs_spark/unsafe/target/scala-2.10/classes...
[error] /home/tgraves/tgravescs_spark/unsafe/src/main/java/org/apache/spark/unsafe/PlatformDependent.java:22: warning: sun.misc.Unsafe is Sun proprietary API and may be removed in a future release
[error] import sun.misc.Unsafe;
[error] ^
[error] /home/tgraves/tgravescs_spark/unsafe/src/main/java/org/apache/spark/unsafe/PlatformDependent.java:26: warning: sun.misc.Unsafe is Sun proprietary API and may be removed in a future release
[error] public static final Unsafe UNSAFE;
[error] ^
[error] /home/tgraves/tgravescs_spark/unsafe/src/main/java/org/apache/spark/unsafe/PlatformDependent.java:43: warning: sun.misc.Unsafe is Sun proprietary API and may be removed in a future release
[error] sun.misc.Unsafe unsafe;
[error] ^
[error] /home/tgraves/tgravescs_spark/unsafe/src/main/java/org/apache/spark/unsafe/PlatformDependent.java:45: warning: sun.misc.Unsafe is Sun proprietary API and may be removed in a future release
[error] Field unsafeField = Unsafe.class.getDeclaredField("theUnsafe");
[error] ^
[error] /home/tgraves/tgravescs_spark/unsafe/src/main/java/org/apache/spark/unsafe/PlatformDependent.java:47: warning: sun.misc.Unsafe is Sun proprietary API and may be removed in a future release
[error] unsafe = (sun.misc.Unsafe) unsafeField.get(null);
[error] ^
[error] /home/tgraves/tgravescs_spark/unsafe/src/main/java/org/apache/spark/unsafe/memory/MemoryBlock.java:53: setMemory(long,long,byte) in sun.misc.Unsafe cannot be applied to (java.lang.Object,long,long,byte)
[error] PlatformDependent.UNSAFE.setMemory(obj, offset, length, (byte) 0);
[error] ^
[error] 1 error
[error] 5 warnings
[error] Compile failed at Apr 30, 2015 2:52:04 PM [0.862s]

Contributor

tgravescs commented Apr 30, 2015

I'm not able to build after this went in. Any special requirements to build this?

[info] Compiling 15 Java sources to /home/tgraves/tgravescs_spark/unsafe/target/scala-2.10/classes...
[error] /home/tgraves/tgravescs_spark/unsafe/src/main/java/org/apache/spark/unsafe/PlatformDependent.java:22: warning: sun.misc.Unsafe is Sun proprietary API and may be removed in a future release
[error] import sun.misc.Unsafe;
[error] ^
[error] /home/tgraves/tgravescs_spark/unsafe/src/main/java/org/apache/spark/unsafe/PlatformDependent.java:26: warning: sun.misc.Unsafe is Sun proprietary API and may be removed in a future release
[error] public static final Unsafe UNSAFE;
[error] ^
[error] /home/tgraves/tgravescs_spark/unsafe/src/main/java/org/apache/spark/unsafe/PlatformDependent.java:43: warning: sun.misc.Unsafe is Sun proprietary API and may be removed in a future release
[error] sun.misc.Unsafe unsafe;
[error] ^
[error] /home/tgraves/tgravescs_spark/unsafe/src/main/java/org/apache/spark/unsafe/PlatformDependent.java:45: warning: sun.misc.Unsafe is Sun proprietary API and may be removed in a future release
[error] Field unsafeField = Unsafe.class.getDeclaredField("theUnsafe");
[error] ^
[error] /home/tgraves/tgravescs_spark/unsafe/src/main/java/org/apache/spark/unsafe/PlatformDependent.java:47: warning: sun.misc.Unsafe is Sun proprietary API and may be removed in a future release
[error] unsafe = (sun.misc.Unsafe) unsafeField.get(null);
[error] ^
[error] /home/tgraves/tgravescs_spark/unsafe/src/main/java/org/apache/spark/unsafe/memory/MemoryBlock.java:53: setMemory(long,long,byte) in sun.misc.Unsafe cannot be applied to (java.lang.Object,long,long,byte)
[error] PlatformDependent.UNSAFE.setMemory(obj, offset, length, (byte) 0);
[error] ^
[error] 1 error
[error] 5 warnings
[error] Compile failed at Apr 30, 2015 2:52:04 PM [0.862s]

@rxin

This comment has been minimized.

Show comment
Hide comment
@rxin

rxin Apr 30, 2015

Contributor

Did you have warning turned on into error?

Contributor

rxin commented Apr 30, 2015

Did you have warning turned on into error?

@tgravescs

This comment has been minimized.

Show comment
Hide comment
@tgravescs

tgravescs Apr 30, 2015

Contributor

not sure what your question is exactly, I think the last one is an error instead of warning. I didn't specifying anything special in build command: mvn -Pyarn -Phadoop-2.4 -DskipTests clean package install

Contributor

tgravescs commented Apr 30, 2015

not sure what your question is exactly, I think the last one is an error instead of warning. I didn't specifying anything special in build command: mvn -Pyarn -Phadoop-2.4 -DskipTests clean package install

@JoshRosen

This comment has been minimized.

Show comment
Hide comment
@JoshRosen

JoshRosen Apr 30, 2015

Contributor

I'm looking into this issue now; these warnings can't be suppressed via the usual annotations mechanism; it requires a special compiler flag. Planning to submit a patch soon; would be great to get your help in testing it.

Contributor

JoshRosen commented Apr 30, 2015

I'm looking into this issue now; these warnings can't be suppressed via the usual annotations mechanism; it requires a special compiler flag. Planning to submit a patch soon; would be great to get your help in testing it.

@tgravescs

This comment has been minimized.

Show comment
Hide comment
@tgravescs

tgravescs Apr 30, 2015

Contributor

Thanks, Also I should have specified, I'm running on a rhel6 box using jdlk6

Contributor

tgravescs commented Apr 30, 2015

Thanks, Also I should have specified, I'm running on a rhel6 box using jdlk6

@vanzin

This comment has been minimized.

Show comment
Hide comment
@vanzin

vanzin Apr 30, 2015

Contributor

Java 6's output:

$ javap sun.misc.Unsafe | grep setMemory
    public native void setMemory(long, long, byte);

So Spark is using an API that doesn't exist there.

Contributor

vanzin commented Apr 30, 2015

Java 6's output:

$ javap sun.misc.Unsafe | grep setMemory
    public native void setMemory(long, long, byte);

So Spark is using an API that doesn't exist there.

@vanzin

This comment has been minimized.

Show comment
Hide comment
@vanzin

vanzin Apr 30, 2015

Contributor

On a side note, if Spark supports java 6, why are all jenkins builds running java 7?

Contributor

vanzin commented Apr 30, 2015

On a side note, if Spark supports java 6, why are all jenkins builds running java 7?

@JoshRosen

This comment has been minimized.

Show comment
Hide comment
@JoshRosen

JoshRosen Apr 30, 2015

Contributor

@tgravescs I've opened #5814 to try to fix the compiler warnings (and error due to Unsafe.setMemory); it would be great if you could test out that commit and see whether it fixes things in your environment.

Contributor

JoshRosen commented Apr 30, 2015

@tgravescs I've opened #5814 to try to fix the compiler warnings (and error due to Unsafe.setMemory); it would be great if you could test out that commit and see whether it fixes things in your environment.

@JoshRosen

This comment has been minimized.

Show comment
Hide comment
@JoshRosen

JoshRosen Apr 30, 2015

Contributor

@vanzin There's an old JIRA for building w/ Java 6 on Jenkins; I've re-opened it: https://issues.apache.org/jira/browse/SPARK-1437

Contributor

JoshRosen commented Apr 30, 2015

@vanzin There's an old JIRA for building w/ Java 6 on Jenkins; I've re-opened it: https://issues.apache.org/jira/browse/SPARK-1437

asfgit pushed a commit that referenced this pull request Apr 30, 2015

[SPARK-7288] Suppress compiler warnings due to use of sun.misc.Unsafe…
…; add facade in front of Unsafe; remove use of Unsafe.setMemory

This patch suppresses compiler warnings due to our use of `sun.misc.Unsafe` (introduced in #5725).  These warnings can only be suppressed via the `-XDignore.symbol.file` javac flag; the `SuppressWarnings` annotation won't work for these.

In order to restrict uses of this compiler flag to the `unsafe` module, I placed a facade in front of `Unsafe` so that other modules won't call it directly. This facade also will also help us to avoid accidental usage of deprecated Unsafe methods or methods that aren't supported in Java 6.

I also removed an unnecessary use of `Unsafe.setMemory`, which isn't present in certain versions of Java 6, and excluded the new `unsafe` module from Javadoc.

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5814 from JoshRosen/unsafe-compiler-warnings-fixes and squashes the following commits:

9e8c483 [Josh Rosen] Exclude new unsafe module from Javadoc
ba75ecf [Josh Rosen] Only apply -XDignore.symbol.file flag in unsafe project.
7403345 [Josh Rosen] Put facade in front of Unsafe.
50230c0 [Josh Rosen] Remove usage of Unsafe.setMemory
96d41c9 [Josh Rosen] Use -XDignore.symbol.file to suppress warnings about sun.misc.Unsafe usage

asfgit pushed a commit that referenced this pull request May 14, 2015

[SPARK-7081] Faster sort-based shuffle path using binary processing c…
…ache-aware sort

This patch introduces a new shuffle manager that enhances the existing sort-based shuffle with a new cache-friendly sort algorithm that operates directly on binary data. The goals of this patch are to lower memory usage and Java object overheads during shuffle and to speed up sorting. It also lays groundwork for follow-up patches that will enable end-to-end processing of serialized records.

The new shuffle manager, `UnsafeShuffleManager`, can be enabled by setting `spark.shuffle.manager=tungsten-sort` in SparkConf.

The new shuffle manager uses directly-managed memory to implement several performance optimizations for certain types of shuffles. In cases where the new performance optimizations cannot be applied, the new shuffle manager delegates to SortShuffleManager to handle those shuffles.

UnsafeShuffleManager's optimizations will apply when _all_ of the following conditions hold:

 - The shuffle dependency specifies no aggregation or output ordering.
 - The shuffle serializer supports relocation of serialized values (this is currently supported
   by KryoSerializer and Spark SQL's custom serializers).
 - The shuffle produces fewer than 16777216 output partitions.
 - No individual record is larger than 128 MB when serialized.

In addition, extra spill-merging optimizations are automatically applied when the shuffle compression codec supports concatenation of serialized streams. This is currently supported by Spark's LZF serializer.

At a high-level, UnsafeShuffleManager's design is similar to Spark's existing SortShuffleManager.  In sort-based shuffle, incoming records are sorted according to their target partition ids, then written to a single map output file. Reducers fetch contiguous regions of this file in order to read their portion of the map output. In cases where the map output data is too large to fit in memory, sorted subsets of the output can are spilled to disk and those on-disk files are merged to produce the final output file.

UnsafeShuffleManager optimizes this process in several ways:

 - Its sort operates on serialized binary data rather than Java objects, which reduces memory consumption and GC overheads. This optimization requires the record serializer to have certain properties to allow serialized records to be re-ordered without requiring deserialization.  See SPARK-4550, where this optimization was first proposed and implemented, for more details.

 - It uses a specialized cache-efficient sorter (UnsafeShuffleExternalSorter) that sorts arrays of compressed record pointers and partition ids. By using only 8 bytes of space per record in the sorting array, this fits more of the array into cache.

 - The spill merging procedure operates on blocks of serialized records that belong to the same partition and does not need to deserialize records during the merge.

 - When the spill compression codec supports concatenation of compressed data, the spill merge simply concatenates the serialized and compressed spill partitions to produce the final output partition.  This allows efficient data copying methods, like NIO's `transferTo`, to be used and avoids the need to allocate decompression or copying buffers during the merge.

The shuffle read path is unchanged.

This patch is similar to [SPARK-4550](http://issues.apache.org/jira/browse/SPARK-4550) / #4450 but uses a slightly different implementation. The `unsafe`-based implementation featured in this patch lays the groundwork for followup patches that will enable sorting to operate on serialized data pages that will be prepared by Spark SQL's new `unsafe` operators (such as the new aggregation operator introduced in #5725).

### Future work

There are several tasks that build upon this patch, which will be left to future work:

- [SPARK-7271](https://issues.apache.org/jira/browse/SPARK-7271) Redesign / extend the shuffle interfaces to accept binary data as input. The goal here is to let us bypass serialization steps in cases where the sort input is produced by an operator that operates directly on binary data.
- Extension / redesign of the `Serializer` API. We can add new methods which allow serializers to determine the size requirements for serializing objects and for serializing objects directly to a specified memory address (similar to how `UnsafeRowConverter` works in Spark SQL).

<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/apache/spark/5868)
<!-- Reviewable:end -->

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5868 from JoshRosen/unsafe-sort and squashes the following commits:

ef0a86e [Josh Rosen] Fix scalastyle errors
7610f2f [Josh Rosen] Add tests for proper cleanup of shuffle data.
d494ffe [Josh Rosen] Fix deserialization of JavaSerializer instances.
52a9981 [Josh Rosen] Fix some bugs in the address packing code.
51812a7 [Josh Rosen] Change shuffle manager sort name to tungsten-sort
4023fa4 [Josh Rosen] Add @Private annotation to some Java classes.
de40b9d [Josh Rosen] More comments to try to explain metrics code
df07699 [Josh Rosen] Attempt to clarify confusing metrics update code
5e189c6 [Josh Rosen] Track time spend closing / flushing files; split TimeTrackingOutputStream into separate file.
d5779c6 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
c2ce78e [Josh Rosen] Fix a missed usage of MAX_PARTITION_ID
e3b8855 [Josh Rosen] Cleanup in UnsafeShuffleWriter
4a2c785 [Josh Rosen] rename 'sort buffer' to 'pointer array'
6276168 [Josh Rosen] Remove ability to disable spilling in UnsafeShuffleExternalSorter.
57312c9 [Josh Rosen] Clarify fileBufferSize units
2d4e4f4 [Josh Rosen] Address some minor comments in UnsafeShuffleExternalSorter.
fdcac08 [Josh Rosen] Guard against overflow when expanding sort buffer.
85da63f [Josh Rosen] Cleanup in UnsafeShuffleSorterIterator.
0ad34da [Josh Rosen] Fix off-by-one in nextInt() call
56781a1 [Josh Rosen] Rename UnsafeShuffleSorter to UnsafeShuffleInMemorySorter
e995d1a [Josh Rosen] Introduce MAX_SHUFFLE_OUTPUT_PARTITIONS.
e58a6b4 [Josh Rosen] Add more tests for PackedRecordPointer encoding.
4f0b770 [Josh Rosen] Attempt to implement proper shuffle write metrics.
d4e6d89 [Josh Rosen] Update to bit shifting constants
69d5899 [Josh Rosen] Remove some unnecessary override vals
8531286 [Josh Rosen] Add tests that automatically trigger spills.
7c953f9 [Josh Rosen] Add test that covers UnsafeShuffleSortDataFormat.swap().
e1855e5 [Josh Rosen] Fix a handful of misc. IntelliJ inspections
39434f9 [Josh Rosen] Avoid integer multiplication overflow in getMemoryUsage (thanks FindBugs!)
1e3ad52 [Josh Rosen] Delete unused ByteBufferOutputStream class.
ea4f85f [Josh Rosen] Roll back an unnecessary change in Spillable.
ae538dc [Josh Rosen] Document UnsafeShuffleManager.
ec6d626 [Josh Rosen] Add notes on maximum # of supported shuffle partitions.
0d4d199 [Josh Rosen] Bump up shuffle.memoryFraction to make tests pass.
b3b1924 [Josh Rosen] Properly implement close() and flush() in DummySerializerInstance.
1ef56c7 [Josh Rosen] Revise compression codec support in merger; test cross product of configurations.
b57c17f [Josh Rosen] Disable some overly-verbose logs that rendered DEBUG useless.
f780fb1 [Josh Rosen] Add test demonstrating which compression codecs support concatenation.
4a01c45 [Josh Rosen] Remove unnecessary log message
27b18b0 [Josh Rosen] That for inserting records AT the max record size.
fcd9a3c [Josh Rosen] Add notes + tests for maximum record / page sizes.
9d1ee7c [Josh Rosen] Fix MiMa excludes for ShuffleWriter change
fd4bb9e [Josh Rosen] Use own ByteBufferOutputStream rather than Kryo's
67d25ba [Josh Rosen] Update Exchange operator's copying logic to account for new shuffle manager
8f5061a [Josh Rosen] Strengthen assertion to check partitioning
01afc74 [Josh Rosen] Actually read data in UnsafeShuffleWriterSuite
1929a74 [Josh Rosen] Update to reflect upstream ShuffleBlockManager -> ShuffleBlockResolver rename.
e8718dd [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
9b7ebed [Josh Rosen] More defensive programming RE: cleaning up spill files and memory after errors
7cd013b [Josh Rosen] Begin refactoring to enable proper tests for spilling.
722849b [Josh Rosen] Add workaround for transferTo() bug in merging code; refactor tests.
9883e30 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
b95e642 [Josh Rosen] Refactor and document logic that decides when to spill.
1ce1300 [Josh Rosen] More minor cleanup
5e8cf75 [Josh Rosen] More minor cleanup
e67f1ea [Josh Rosen] Remove upper type bound in ShuffleWriter interface.
cfe0ec4 [Josh Rosen] Address a number of minor review comments:
8a6fe52 [Josh Rosen] Rename UnsafeShuffleSpillWriter to UnsafeShuffleExternalSorter
11feeb6 [Josh Rosen] Update TODOs related to shuffle write metrics.
b674412 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
aaea17b [Josh Rosen] Add comments to UnsafeShuffleSpillWriter.
4f70141 [Josh Rosen] Fix merging; now passes UnsafeShuffleSuite tests.
133c8c9 [Josh Rosen] WIP towards testing UnsafeShuffleWriter.
f480fb2 [Josh Rosen] WIP in mega-refactoring towards shuffle-specific sort.
57f1ec0 [Josh Rosen] WIP towards packed record pointers for use in optimized shuffle sort.
69232fd [Josh Rosen] Enable compressible address encoding for off-heap mode.
7ee918e [Josh Rosen] Re-order imports in tests
3aeaff7 [Josh Rosen] More refactoring and cleanup; begin cleaning iterator interfaces
3490512 [Josh Rosen] Misc. cleanup
f156a8f [Josh Rosen] Hacky metrics integration; refactor some interfaces.
2776aca [Josh Rosen] First passing test for ExternalSorter.
5e100b2 [Josh Rosen] Super-messy WIP on external sort
595923a [Josh Rosen] Remove some unused variables.
8958584 [Josh Rosen] Fix bug in calculating free space in current page.
f17fa8f [Josh Rosen] Add missing newline
c2fca17 [Josh Rosen] Small refactoring of SerializerPropertiesSuite to enable test re-use:
b8a09fe [Josh Rosen] Back out accidental log4j.properties change
bfc12d3 [Josh Rosen] Add tests for serializer relocation property.
240864c [Josh Rosen] Remove PrefixComputer and require prefix to be specified as part of insert()
1433b42 [Josh Rosen] Store record length as int instead of long.
026b497 [Josh Rosen] Re-use a buffer in UnsafeShuffleWriter
0748458 [Josh Rosen] Port UnsafeShuffleWriter to Java.
87e721b [Josh Rosen] Renaming and comments
d3cc310 [Josh Rosen] Flag that SparkSqlSerializer2 supports relocation
e2d96ca [Josh Rosen] Expand serializer API and use new function to help control when new UnsafeShuffle path is used.
e267cee [Josh Rosen] Fix compilation of UnsafeSorterSuite
9c6cf58 [Josh Rosen] Refactor to use DiskBlockObjectWriter.
253f13e [Josh Rosen] More cleanup
8e3ec20 [Josh Rosen] Begin code cleanup.
4d2f5e1 [Josh Rosen] WIP
3db12de [Josh Rosen] Minor simplification and sanity checks in UnsafeSorter
767d3ca [Josh Rosen] Fix invalid range in UnsafeSorter.
e900152 [Josh Rosen] Add test for empty iterator in UnsafeSorter
57a4ea0 [Josh Rosen] Make initialSize configurable in UnsafeSorter
abf7bfe [Josh Rosen] Add basic test case.
81d52c5 [Josh Rosen] WIP on UnsafeSorter

(cherry picked from commit 73bed40)
Signed-off-by: Reynold Xin <rxin@databricks.com>

asfgit pushed a commit that referenced this pull request May 14, 2015

[SPARK-7081] Faster sort-based shuffle path using binary processing c…
…ache-aware sort

This patch introduces a new shuffle manager that enhances the existing sort-based shuffle with a new cache-friendly sort algorithm that operates directly on binary data. The goals of this patch are to lower memory usage and Java object overheads during shuffle and to speed up sorting. It also lays groundwork for follow-up patches that will enable end-to-end processing of serialized records.

The new shuffle manager, `UnsafeShuffleManager`, can be enabled by setting `spark.shuffle.manager=tungsten-sort` in SparkConf.

The new shuffle manager uses directly-managed memory to implement several performance optimizations for certain types of shuffles. In cases where the new performance optimizations cannot be applied, the new shuffle manager delegates to SortShuffleManager to handle those shuffles.

UnsafeShuffleManager's optimizations will apply when _all_ of the following conditions hold:

 - The shuffle dependency specifies no aggregation or output ordering.
 - The shuffle serializer supports relocation of serialized values (this is currently supported
   by KryoSerializer and Spark SQL's custom serializers).
 - The shuffle produces fewer than 16777216 output partitions.
 - No individual record is larger than 128 MB when serialized.

In addition, extra spill-merging optimizations are automatically applied when the shuffle compression codec supports concatenation of serialized streams. This is currently supported by Spark's LZF serializer.

At a high-level, UnsafeShuffleManager's design is similar to Spark's existing SortShuffleManager.  In sort-based shuffle, incoming records are sorted according to their target partition ids, then written to a single map output file. Reducers fetch contiguous regions of this file in order to read their portion of the map output. In cases where the map output data is too large to fit in memory, sorted subsets of the output can are spilled to disk and those on-disk files are merged to produce the final output file.

UnsafeShuffleManager optimizes this process in several ways:

 - Its sort operates on serialized binary data rather than Java objects, which reduces memory consumption and GC overheads. This optimization requires the record serializer to have certain properties to allow serialized records to be re-ordered without requiring deserialization.  See SPARK-4550, where this optimization was first proposed and implemented, for more details.

 - It uses a specialized cache-efficient sorter (UnsafeShuffleExternalSorter) that sorts arrays of compressed record pointers and partition ids. By using only 8 bytes of space per record in the sorting array, this fits more of the array into cache.

 - The spill merging procedure operates on blocks of serialized records that belong to the same partition and does not need to deserialize records during the merge.

 - When the spill compression codec supports concatenation of compressed data, the spill merge simply concatenates the serialized and compressed spill partitions to produce the final output partition.  This allows efficient data copying methods, like NIO's `transferTo`, to be used and avoids the need to allocate decompression or copying buffers during the merge.

The shuffle read path is unchanged.

This patch is similar to [SPARK-4550](http://issues.apache.org/jira/browse/SPARK-4550) / #4450 but uses a slightly different implementation. The `unsafe`-based implementation featured in this patch lays the groundwork for followup patches that will enable sorting to operate on serialized data pages that will be prepared by Spark SQL's new `unsafe` operators (such as the new aggregation operator introduced in #5725).

### Future work

There are several tasks that build upon this patch, which will be left to future work:

- [SPARK-7271](https://issues.apache.org/jira/browse/SPARK-7271) Redesign / extend the shuffle interfaces to accept binary data as input. The goal here is to let us bypass serialization steps in cases where the sort input is produced by an operator that operates directly on binary data.
- Extension / redesign of the `Serializer` API. We can add new methods which allow serializers to determine the size requirements for serializing objects and for serializing objects directly to a specified memory address (similar to how `UnsafeRowConverter` works in Spark SQL).

<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/apache/spark/5868)
<!-- Reviewable:end -->

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5868 from JoshRosen/unsafe-sort and squashes the following commits:

ef0a86e [Josh Rosen] Fix scalastyle errors
7610f2f [Josh Rosen] Add tests for proper cleanup of shuffle data.
d494ffe [Josh Rosen] Fix deserialization of JavaSerializer instances.
52a9981 [Josh Rosen] Fix some bugs in the address packing code.
51812a7 [Josh Rosen] Change shuffle manager sort name to tungsten-sort
4023fa4 [Josh Rosen] Add @Private annotation to some Java classes.
de40b9d [Josh Rosen] More comments to try to explain metrics code
df07699 [Josh Rosen] Attempt to clarify confusing metrics update code
5e189c6 [Josh Rosen] Track time spend closing / flushing files; split TimeTrackingOutputStream into separate file.
d5779c6 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
c2ce78e [Josh Rosen] Fix a missed usage of MAX_PARTITION_ID
e3b8855 [Josh Rosen] Cleanup in UnsafeShuffleWriter
4a2c785 [Josh Rosen] rename 'sort buffer' to 'pointer array'
6276168 [Josh Rosen] Remove ability to disable spilling in UnsafeShuffleExternalSorter.
57312c9 [Josh Rosen] Clarify fileBufferSize units
2d4e4f4 [Josh Rosen] Address some minor comments in UnsafeShuffleExternalSorter.
fdcac08 [Josh Rosen] Guard against overflow when expanding sort buffer.
85da63f [Josh Rosen] Cleanup in UnsafeShuffleSorterIterator.
0ad34da [Josh Rosen] Fix off-by-one in nextInt() call
56781a1 [Josh Rosen] Rename UnsafeShuffleSorter to UnsafeShuffleInMemorySorter
e995d1a [Josh Rosen] Introduce MAX_SHUFFLE_OUTPUT_PARTITIONS.
e58a6b4 [Josh Rosen] Add more tests for PackedRecordPointer encoding.
4f0b770 [Josh Rosen] Attempt to implement proper shuffle write metrics.
d4e6d89 [Josh Rosen] Update to bit shifting constants
69d5899 [Josh Rosen] Remove some unnecessary override vals
8531286 [Josh Rosen] Add tests that automatically trigger spills.
7c953f9 [Josh Rosen] Add test that covers UnsafeShuffleSortDataFormat.swap().
e1855e5 [Josh Rosen] Fix a handful of misc. IntelliJ inspections
39434f9 [Josh Rosen] Avoid integer multiplication overflow in getMemoryUsage (thanks FindBugs!)
1e3ad52 [Josh Rosen] Delete unused ByteBufferOutputStream class.
ea4f85f [Josh Rosen] Roll back an unnecessary change in Spillable.
ae538dc [Josh Rosen] Document UnsafeShuffleManager.
ec6d626 [Josh Rosen] Add notes on maximum # of supported shuffle partitions.
0d4d199 [Josh Rosen] Bump up shuffle.memoryFraction to make tests pass.
b3b1924 [Josh Rosen] Properly implement close() and flush() in DummySerializerInstance.
1ef56c7 [Josh Rosen] Revise compression codec support in merger; test cross product of configurations.
b57c17f [Josh Rosen] Disable some overly-verbose logs that rendered DEBUG useless.
f780fb1 [Josh Rosen] Add test demonstrating which compression codecs support concatenation.
4a01c45 [Josh Rosen] Remove unnecessary log message
27b18b0 [Josh Rosen] That for inserting records AT the max record size.
fcd9a3c [Josh Rosen] Add notes + tests for maximum record / page sizes.
9d1ee7c [Josh Rosen] Fix MiMa excludes for ShuffleWriter change
fd4bb9e [Josh Rosen] Use own ByteBufferOutputStream rather than Kryo's
67d25ba [Josh Rosen] Update Exchange operator's copying logic to account for new shuffle manager
8f5061a [Josh Rosen] Strengthen assertion to check partitioning
01afc74 [Josh Rosen] Actually read data in UnsafeShuffleWriterSuite
1929a74 [Josh Rosen] Update to reflect upstream ShuffleBlockManager -> ShuffleBlockResolver rename.
e8718dd [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
9b7ebed [Josh Rosen] More defensive programming RE: cleaning up spill files and memory after errors
7cd013b [Josh Rosen] Begin refactoring to enable proper tests for spilling.
722849b [Josh Rosen] Add workaround for transferTo() bug in merging code; refactor tests.
9883e30 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
b95e642 [Josh Rosen] Refactor and document logic that decides when to spill.
1ce1300 [Josh Rosen] More minor cleanup
5e8cf75 [Josh Rosen] More minor cleanup
e67f1ea [Josh Rosen] Remove upper type bound in ShuffleWriter interface.
cfe0ec4 [Josh Rosen] Address a number of minor review comments:
8a6fe52 [Josh Rosen] Rename UnsafeShuffleSpillWriter to UnsafeShuffleExternalSorter
11feeb6 [Josh Rosen] Update TODOs related to shuffle write metrics.
b674412 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
aaea17b [Josh Rosen] Add comments to UnsafeShuffleSpillWriter.
4f70141 [Josh Rosen] Fix merging; now passes UnsafeShuffleSuite tests.
133c8c9 [Josh Rosen] WIP towards testing UnsafeShuffleWriter.
f480fb2 [Josh Rosen] WIP in mega-refactoring towards shuffle-specific sort.
57f1ec0 [Josh Rosen] WIP towards packed record pointers for use in optimized shuffle sort.
69232fd [Josh Rosen] Enable compressible address encoding for off-heap mode.
7ee918e [Josh Rosen] Re-order imports in tests
3aeaff7 [Josh Rosen] More refactoring and cleanup; begin cleaning iterator interfaces
3490512 [Josh Rosen] Misc. cleanup
f156a8f [Josh Rosen] Hacky metrics integration; refactor some interfaces.
2776aca [Josh Rosen] First passing test for ExternalSorter.
5e100b2 [Josh Rosen] Super-messy WIP on external sort
595923a [Josh Rosen] Remove some unused variables.
8958584 [Josh Rosen] Fix bug in calculating free space in current page.
f17fa8f [Josh Rosen] Add missing newline
c2fca17 [Josh Rosen] Small refactoring of SerializerPropertiesSuite to enable test re-use:
b8a09fe [Josh Rosen] Back out accidental log4j.properties change
bfc12d3 [Josh Rosen] Add tests for serializer relocation property.
240864c [Josh Rosen] Remove PrefixComputer and require prefix to be specified as part of insert()
1433b42 [Josh Rosen] Store record length as int instead of long.
026b497 [Josh Rosen] Re-use a buffer in UnsafeShuffleWriter
0748458 [Josh Rosen] Port UnsafeShuffleWriter to Java.
87e721b [Josh Rosen] Renaming and comments
d3cc310 [Josh Rosen] Flag that SparkSqlSerializer2 supports relocation
e2d96ca [Josh Rosen] Expand serializer API and use new function to help control when new UnsafeShuffle path is used.
e267cee [Josh Rosen] Fix compilation of UnsafeSorterSuite
9c6cf58 [Josh Rosen] Refactor to use DiskBlockObjectWriter.
253f13e [Josh Rosen] More cleanup
8e3ec20 [Josh Rosen] Begin code cleanup.
4d2f5e1 [Josh Rosen] WIP
3db12de [Josh Rosen] Minor simplification and sanity checks in UnsafeSorter
767d3ca [Josh Rosen] Fix invalid range in UnsafeSorter.
e900152 [Josh Rosen] Add test for empty iterator in UnsafeSorter
57a4ea0 [Josh Rosen] Make initialSize configurable in UnsafeSorter
abf7bfe [Josh Rosen] Add basic test case.
81d52c5 [Josh Rosen] WIP on UnsafeSorter
* retrieve the base object.
* <p>
* This allows us to address 8192 pages. In on-heap mode, the maximum page size is limited by the
* maximum size of a long[] array, allowing us to address 8192 * 2^32 * 8 bytes, which is

This comment has been minimized.

@cloud-fan

cloud-fan May 21, 2015

Contributor

Should this be 8192 * 2^31 * 8 bytes? The length of array is up to Integer.MAX_VALUE.

@cloud-fan

cloud-fan May 21, 2015

Contributor

Should this be 8192 * 2^31 * 8 bytes? The length of array is up to Integer.MAX_VALUE.

jeanlyn added a commit to jeanlyn/spark that referenced this pull request May 28, 2015

[SPARK-7076][SPARK-7077][SPARK-7080][SQL] Use managed memory for aggr…
…egations

This patch adds managed-memory-based aggregation to Spark SQL / DataFrames. Instead of working with Java objects, this new aggregation path uses `sun.misc.Unsafe` to manipulate raw memory.  This reduces the memory footprint for aggregations, resulting in fewer spills, OutOfMemoryErrors, and garbage collection pauses.  As a result, this allows for higher memory utilization.  It can also result in better cache locality since objects will be stored closer together in memory.

This feature can be eanbled by setting `spark.sql.unsafe.enabled=true`.  For now, this feature is only supported when codegen is enabled and only supports aggregations for which the grouping columns are primitive numeric types or strings and aggregated values are numeric.

### Managing memory with sun.misc.Unsafe

This patch supports both on- and off-heap managed memory.

- In on-heap mode, memory addresses are identified by the combination of a base Object and an offset within that object.
- In off-heap mode, memory is addressed directly with 64-bit long addresses.

To support both modes, functions that manipulate memory accept both `baseObject` and `baseOffset` fields.  In off-heap mode, we simply pass `null` as `baseObject`.

We allocate memory in large chunks, so memory fragmentation and allocation speed are not significant bottlenecks.

By default, we use on-heap mode.  To enable off-heap mode, set `spark.unsafe.offHeap=true`.

To track allocated memory, this patch extends `SparkEnv` with an `ExecutorMemoryManager` and supplies each `TaskContext` with a `TaskMemoryManager`.  These classes work together to track allocations and detect memory leaks.

### Compact tuple format

This patch introduces `UnsafeRow`, a compact row layout.  In this format, each tuple has three parts: a null bit set, fixed length values, and variable-length values:

![image](https://cloud.githubusercontent.com/assets/50748/7328538/2fdb65ce-ea8b-11e4-9743-6c0f02bb7d1f.png)

- Rows are always 8-byte word aligned (so their sizes will always be a multiple of 8 bytes)
- The bit set is used for null tracking:
	- Position _i_ is set if and only if field _i_ is null
	- The bit set is aligned to an 8-byte word boundary.
- Every field appears as an 8-byte word in the fixed-length values part:
	- If a field is null, we zero out the values.
	- If a field is variable-length, the word stores a relative offset (w.r.t. the base of the tuple) that points to the beginning of the field's data in the variable-length part.
- Each variable-length data type can have its own encoding:
	- For strings, the first word stores the length of the string and is followed by UTF-8 encoded bytes.  If necessary, the end of the string is padded with empty bytes in order to ensure word-alignment.

For example, a tuple that consists 3 fields of type (int, string, string), with value (null, “data”, “bricks”) would look like this:

![image](https://cloud.githubusercontent.com/assets/50748/7328526/1e21959c-ea8b-11e4-9a28-a4350fe4a7b5.png)

This format allows us to compare tuples for equality by directly comparing their raw bytes.  This also enables fast hashing of tuples.

### Hash map for performing aggregations

This patch introduces `UnsafeFixedWidthAggregationMap`, a hash map for performing aggregations where the aggregation result columns are fixed-with.  This map's keys and values are `Row` objects. `UnsafeFixedWidthAggregationMap` is implemented on top of `BytesToBytesMap`, an append-only map which supports byte-array keys and values.

`BytesToBytesMap` stores pointers to key and value tuples.  For each record with a new key, we copy the key and create the aggregation value buffer for that key and put them in a buffer. The hash table then simply stores pointers to the key and value. For each record with an existing key, we simply run the aggregation function to update the values in place.

This map is implemented using open hashing with triangular sequence probing.  Each entry stores two words in a long array: the first word stores the address of the key and the second word stores the relative offset from the key tuple to the value tuple, as well as the key's 32-bit hashcode.  By storing the full hashcode, we reduce the number of equality checks that need to be performed to handle position collisions ()since the chance of hashcode collision is much lower than position collision).

`UnsafeFixedWidthAggregationMap` allows regular Spark SQL `Row` objects to be used when probing the map.  Internally, it encodes these rows into `UnsafeRow` format using `UnsafeRowConverter`.  This conversion has a small overhead that can be eliminated in the future once we use UnsafeRows in other operators.

<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/apache/spark/5725)
<!-- Reviewable:end -->

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5725 from JoshRosen/unsafe and squashes the following commits:

eeee512 [Josh Rosen] Add converters for Null, Boolean, Byte, and Short columns.
81f34f8 [Josh Rosen] Follow 'place children last' convention for GeneratedAggregate
1bc36cc [Josh Rosen] Refactor UnsafeRowConverter to avoid unnecessary boxing.
017b2dc [Josh Rosen] Remove BytesToBytesMap.finalize()
50e9671 [Josh Rosen] Throw memory leak warning even in case of error; add warning about code duplication
70a39e4 [Josh Rosen] Split MemoryManager into ExecutorMemoryManager and TaskMemoryManager:
6e4b192 [Josh Rosen] Remove an unused method from ByteArrayMethods.
de5e001 [Josh Rosen] Fix debug vs. trace in logging message.
a19e066 [Josh Rosen] Rename unsafe Java test suites to match Scala test naming convention.
78a5b84 [Josh Rosen] Add logging to MemoryManager
ce3c565 [Josh Rosen] More comments, formatting, and code cleanup.
529e571 [Josh Rosen] Measure timeSpentResizing in nanoseconds instead of milliseconds.
3ca84b2 [Josh Rosen] Only zero the used portion of groupingKeyConversionScratchSpace
162caf7 [Josh Rosen] Fix test compilation
b45f070 [Josh Rosen] Don't redundantly store the offset from key to value, since we can compute this from the key size.
a8e4a3f [Josh Rosen] Introduce MemoryManager interface; add to SparkEnv.
0925847 [Josh Rosen] Disable MiMa checks for new unsafe module
cde4132 [Josh Rosen] Add missing pom.xml
9c19fc0 [Josh Rosen] Add configuration options for heap vs. offheap
6ffdaa1 [Josh Rosen] Null handling improvements in UnsafeRow.
31eaabc [Josh Rosen] Lots of TODO and doc cleanup.
a95291e [Josh Rosen] Cleanups to string handling code
afe8dca [Josh Rosen] Some Javadoc cleanup
f3dcbfe [Josh Rosen] More mod replacement
854201a [Josh Rosen] Import and comment cleanup
06e929d [Josh Rosen] More warning cleanup
ef6b3d3 [Josh Rosen] Fix a bunch of FindBugs and IntelliJ inspections
29a7575 [Josh Rosen] Remove debug logging
49aed30 [Josh Rosen] More long -> int conversion.
b26f1d3 [Josh Rosen] Fix bug in murmur hash implementation.
765243d [Josh Rosen] Enable optional performance metrics for hash map.
23a440a [Josh Rosen] Bump up default hash map size
628f936 [Josh Rosen] Use ints intead of longs for indexing.
92d5a06 [Josh Rosen] Address a number of minor code review comments.
1f4b716 [Josh Rosen] Merge Unsafe code into the regular GeneratedAggregate, guarded by a configuration flag; integrate planner support and re-enable all tests.
d85eeff [Josh Rosen] Add basic sanity test for UnsafeFixedWidthAggregationMap
bade966 [Josh Rosen] Comment update (bumping to refresh GitHub cache...)
b3eaccd [Josh Rosen] Extract aggregation map into its own class.
d2bb986 [Josh Rosen] Update to implement new Row methods added upstream
58ac393 [Josh Rosen] Use UNSAFE allocator in GeneratedAggregate (TODO: make this configurable)
7df6008 [Josh Rosen] Optimizations related to zeroing out memory:
c1b3813 [Josh Rosen] Fix bug in UnsafeMemoryAllocator.free():
738fa33 [Josh Rosen] Add feature flag to guard UnsafeGeneratedAggregate
c55bf66 [Josh Rosen] Free buffer once iterator has been fully consumed.
62ab054 [Josh Rosen] Optimize for fact that get() is only called on String columns.
c7f0b56 [Josh Rosen] Reuse UnsafeRow pointer in UnsafeRowConverter
ae39694 [Josh Rosen] Add finalizer as "cleanup method of last resort"
c754ae1 [Josh Rosen] Now that the store*() contract has been stregthened, we can remove an extra lookup
f764d13 [Josh Rosen] Simplify address + length calculation in Location.
079f1bf [Josh Rosen] Some clarification of the BytesToBytesMap.lookup() / set() contract.
1a483c5 [Josh Rosen] First version that passes some aggregation tests:
fc4c3a8 [Josh Rosen] Sketch how the converters will be used in UnsafeGeneratedAggregate
53ba9b7 [Josh Rosen] Start prototyping Java Row -> UnsafeRow converters
1ff814d [Josh Rosen] Add reminder to free memory on iterator completion
8a8f9df [Josh Rosen] Add skeleton for GeneratedAggregate integration.
5d55cef [Josh Rosen] Add skeleton for Row implementation.
f03e9c1 [Josh Rosen] Play around with Unsafe implementations of more string methods.
ab68e08 [Josh Rosen] Begin merging the UTF8String implementations.
480a74a [Josh Rosen] Initial import of code from Databricks unsafe utils repo.

jeanlyn added a commit to jeanlyn/spark that referenced this pull request May 28, 2015

[SPARK-7288] Suppress compiler warnings due to use of sun.misc.Unsafe…
…; add facade in front of Unsafe; remove use of Unsafe.setMemory

This patch suppresses compiler warnings due to our use of `sun.misc.Unsafe` (introduced in #5725).  These warnings can only be suppressed via the `-XDignore.symbol.file` javac flag; the `SuppressWarnings` annotation won't work for these.

In order to restrict uses of this compiler flag to the `unsafe` module, I placed a facade in front of `Unsafe` so that other modules won't call it directly. This facade also will also help us to avoid accidental usage of deprecated Unsafe methods or methods that aren't supported in Java 6.

I also removed an unnecessary use of `Unsafe.setMemory`, which isn't present in certain versions of Java 6, and excluded the new `unsafe` module from Javadoc.

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5814 from JoshRosen/unsafe-compiler-warnings-fixes and squashes the following commits:

9e8c483 [Josh Rosen] Exclude new unsafe module from Javadoc
ba75ecf [Josh Rosen] Only apply -XDignore.symbol.file flag in unsafe project.
7403345 [Josh Rosen] Put facade in front of Unsafe.
50230c0 [Josh Rosen] Remove usage of Unsafe.setMemory
96d41c9 [Josh Rosen] Use -XDignore.symbol.file to suppress warnings about sun.misc.Unsafe usage

jeanlyn added a commit to jeanlyn/spark that referenced this pull request May 28, 2015

[SPARK-7081] Faster sort-based shuffle path using binary processing c…
…ache-aware sort

This patch introduces a new shuffle manager that enhances the existing sort-based shuffle with a new cache-friendly sort algorithm that operates directly on binary data. The goals of this patch are to lower memory usage and Java object overheads during shuffle and to speed up sorting. It also lays groundwork for follow-up patches that will enable end-to-end processing of serialized records.

The new shuffle manager, `UnsafeShuffleManager`, can be enabled by setting `spark.shuffle.manager=tungsten-sort` in SparkConf.

The new shuffle manager uses directly-managed memory to implement several performance optimizations for certain types of shuffles. In cases where the new performance optimizations cannot be applied, the new shuffle manager delegates to SortShuffleManager to handle those shuffles.

UnsafeShuffleManager's optimizations will apply when _all_ of the following conditions hold:

 - The shuffle dependency specifies no aggregation or output ordering.
 - The shuffle serializer supports relocation of serialized values (this is currently supported
   by KryoSerializer and Spark SQL's custom serializers).
 - The shuffle produces fewer than 16777216 output partitions.
 - No individual record is larger than 128 MB when serialized.

In addition, extra spill-merging optimizations are automatically applied when the shuffle compression codec supports concatenation of serialized streams. This is currently supported by Spark's LZF serializer.

At a high-level, UnsafeShuffleManager's design is similar to Spark's existing SortShuffleManager.  In sort-based shuffle, incoming records are sorted according to their target partition ids, then written to a single map output file. Reducers fetch contiguous regions of this file in order to read their portion of the map output. In cases where the map output data is too large to fit in memory, sorted subsets of the output can are spilled to disk and those on-disk files are merged to produce the final output file.

UnsafeShuffleManager optimizes this process in several ways:

 - Its sort operates on serialized binary data rather than Java objects, which reduces memory consumption and GC overheads. This optimization requires the record serializer to have certain properties to allow serialized records to be re-ordered without requiring deserialization.  See SPARK-4550, where this optimization was first proposed and implemented, for more details.

 - It uses a specialized cache-efficient sorter (UnsafeShuffleExternalSorter) that sorts arrays of compressed record pointers and partition ids. By using only 8 bytes of space per record in the sorting array, this fits more of the array into cache.

 - The spill merging procedure operates on blocks of serialized records that belong to the same partition and does not need to deserialize records during the merge.

 - When the spill compression codec supports concatenation of compressed data, the spill merge simply concatenates the serialized and compressed spill partitions to produce the final output partition.  This allows efficient data copying methods, like NIO's `transferTo`, to be used and avoids the need to allocate decompression or copying buffers during the merge.

The shuffle read path is unchanged.

This patch is similar to [SPARK-4550](http://issues.apache.org/jira/browse/SPARK-4550) / #4450 but uses a slightly different implementation. The `unsafe`-based implementation featured in this patch lays the groundwork for followup patches that will enable sorting to operate on serialized data pages that will be prepared by Spark SQL's new `unsafe` operators (such as the new aggregation operator introduced in #5725).

### Future work

There are several tasks that build upon this patch, which will be left to future work:

- [SPARK-7271](https://issues.apache.org/jira/browse/SPARK-7271) Redesign / extend the shuffle interfaces to accept binary data as input. The goal here is to let us bypass serialization steps in cases where the sort input is produced by an operator that operates directly on binary data.
- Extension / redesign of the `Serializer` API. We can add new methods which allow serializers to determine the size requirements for serializing objects and for serializing objects directly to a specified memory address (similar to how `UnsafeRowConverter` works in Spark SQL).

<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/apache/spark/5868)
<!-- Reviewable:end -->

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5868 from JoshRosen/unsafe-sort and squashes the following commits:

ef0a86e [Josh Rosen] Fix scalastyle errors
7610f2f [Josh Rosen] Add tests for proper cleanup of shuffle data.
d494ffe [Josh Rosen] Fix deserialization of JavaSerializer instances.
52a9981 [Josh Rosen] Fix some bugs in the address packing code.
51812a7 [Josh Rosen] Change shuffle manager sort name to tungsten-sort
4023fa4 [Josh Rosen] Add @Private annotation to some Java classes.
de40b9d [Josh Rosen] More comments to try to explain metrics code
df07699 [Josh Rosen] Attempt to clarify confusing metrics update code
5e189c6 [Josh Rosen] Track time spend closing / flushing files; split TimeTrackingOutputStream into separate file.
d5779c6 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
c2ce78e [Josh Rosen] Fix a missed usage of MAX_PARTITION_ID
e3b8855 [Josh Rosen] Cleanup in UnsafeShuffleWriter
4a2c785 [Josh Rosen] rename 'sort buffer' to 'pointer array'
6276168 [Josh Rosen] Remove ability to disable spilling in UnsafeShuffleExternalSorter.
57312c9 [Josh Rosen] Clarify fileBufferSize units
2d4e4f4 [Josh Rosen] Address some minor comments in UnsafeShuffleExternalSorter.
fdcac08 [Josh Rosen] Guard against overflow when expanding sort buffer.
85da63f [Josh Rosen] Cleanup in UnsafeShuffleSorterIterator.
0ad34da [Josh Rosen] Fix off-by-one in nextInt() call
56781a1 [Josh Rosen] Rename UnsafeShuffleSorter to UnsafeShuffleInMemorySorter
e995d1a [Josh Rosen] Introduce MAX_SHUFFLE_OUTPUT_PARTITIONS.
e58a6b4 [Josh Rosen] Add more tests for PackedRecordPointer encoding.
4f0b770 [Josh Rosen] Attempt to implement proper shuffle write metrics.
d4e6d89 [Josh Rosen] Update to bit shifting constants
69d5899 [Josh Rosen] Remove some unnecessary override vals
8531286 [Josh Rosen] Add tests that automatically trigger spills.
7c953f9 [Josh Rosen] Add test that covers UnsafeShuffleSortDataFormat.swap().
e1855e5 [Josh Rosen] Fix a handful of misc. IntelliJ inspections
39434f9 [Josh Rosen] Avoid integer multiplication overflow in getMemoryUsage (thanks FindBugs!)
1e3ad52 [Josh Rosen] Delete unused ByteBufferOutputStream class.
ea4f85f [Josh Rosen] Roll back an unnecessary change in Spillable.
ae538dc [Josh Rosen] Document UnsafeShuffleManager.
ec6d626 [Josh Rosen] Add notes on maximum # of supported shuffle partitions.
0d4d199 [Josh Rosen] Bump up shuffle.memoryFraction to make tests pass.
b3b1924 [Josh Rosen] Properly implement close() and flush() in DummySerializerInstance.
1ef56c7 [Josh Rosen] Revise compression codec support in merger; test cross product of configurations.
b57c17f [Josh Rosen] Disable some overly-verbose logs that rendered DEBUG useless.
f780fb1 [Josh Rosen] Add test demonstrating which compression codecs support concatenation.
4a01c45 [Josh Rosen] Remove unnecessary log message
27b18b0 [Josh Rosen] That for inserting records AT the max record size.
fcd9a3c [Josh Rosen] Add notes + tests for maximum record / page sizes.
9d1ee7c [Josh Rosen] Fix MiMa excludes for ShuffleWriter change
fd4bb9e [Josh Rosen] Use own ByteBufferOutputStream rather than Kryo's
67d25ba [Josh Rosen] Update Exchange operator's copying logic to account for new shuffle manager
8f5061a [Josh Rosen] Strengthen assertion to check partitioning
01afc74 [Josh Rosen] Actually read data in UnsafeShuffleWriterSuite
1929a74 [Josh Rosen] Update to reflect upstream ShuffleBlockManager -> ShuffleBlockResolver rename.
e8718dd [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
9b7ebed [Josh Rosen] More defensive programming RE: cleaning up spill files and memory after errors
7cd013b [Josh Rosen] Begin refactoring to enable proper tests for spilling.
722849b [Josh Rosen] Add workaround for transferTo() bug in merging code; refactor tests.
9883e30 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
b95e642 [Josh Rosen] Refactor and document logic that decides when to spill.
1ce1300 [Josh Rosen] More minor cleanup
5e8cf75 [Josh Rosen] More minor cleanup
e67f1ea [Josh Rosen] Remove upper type bound in ShuffleWriter interface.
cfe0ec4 [Josh Rosen] Address a number of minor review comments:
8a6fe52 [Josh Rosen] Rename UnsafeShuffleSpillWriter to UnsafeShuffleExternalSorter
11feeb6 [Josh Rosen] Update TODOs related to shuffle write metrics.
b674412 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
aaea17b [Josh Rosen] Add comments to UnsafeShuffleSpillWriter.
4f70141 [Josh Rosen] Fix merging; now passes UnsafeShuffleSuite tests.
133c8c9 [Josh Rosen] WIP towards testing UnsafeShuffleWriter.
f480fb2 [Josh Rosen] WIP in mega-refactoring towards shuffle-specific sort.
57f1ec0 [Josh Rosen] WIP towards packed record pointers for use in optimized shuffle sort.
69232fd [Josh Rosen] Enable compressible address encoding for off-heap mode.
7ee918e [Josh Rosen] Re-order imports in tests
3aeaff7 [Josh Rosen] More refactoring and cleanup; begin cleaning iterator interfaces
3490512 [Josh Rosen] Misc. cleanup
f156a8f [Josh Rosen] Hacky metrics integration; refactor some interfaces.
2776aca [Josh Rosen] First passing test for ExternalSorter.
5e100b2 [Josh Rosen] Super-messy WIP on external sort
595923a [Josh Rosen] Remove some unused variables.
8958584 [Josh Rosen] Fix bug in calculating free space in current page.
f17fa8f [Josh Rosen] Add missing newline
c2fca17 [Josh Rosen] Small refactoring of SerializerPropertiesSuite to enable test re-use:
b8a09fe [Josh Rosen] Back out accidental log4j.properties change
bfc12d3 [Josh Rosen] Add tests for serializer relocation property.
240864c [Josh Rosen] Remove PrefixComputer and require prefix to be specified as part of insert()
1433b42 [Josh Rosen] Store record length as int instead of long.
026b497 [Josh Rosen] Re-use a buffer in UnsafeShuffleWriter
0748458 [Josh Rosen] Port UnsafeShuffleWriter to Java.
87e721b [Josh Rosen] Renaming and comments
d3cc310 [Josh Rosen] Flag that SparkSqlSerializer2 supports relocation
e2d96ca [Josh Rosen] Expand serializer API and use new function to help control when new UnsafeShuffle path is used.
e267cee [Josh Rosen] Fix compilation of UnsafeSorterSuite
9c6cf58 [Josh Rosen] Refactor to use DiskBlockObjectWriter.
253f13e [Josh Rosen] More cleanup
8e3ec20 [Josh Rosen] Begin code cleanup.
4d2f5e1 [Josh Rosen] WIP
3db12de [Josh Rosen] Minor simplification and sanity checks in UnsafeSorter
767d3ca [Josh Rosen] Fix invalid range in UnsafeSorter.
e900152 [Josh Rosen] Add test for empty iterator in UnsafeSorter
57a4ea0 [Josh Rosen] Make initialSize configurable in UnsafeSorter
abf7bfe [Josh Rosen] Add basic test case.
81d52c5 [Josh Rosen] WIP on UnsafeSorter

jeanlyn added a commit to jeanlyn/spark that referenced this pull request Jun 12, 2015

[SPARK-7076][SPARK-7077][SPARK-7080][SQL] Use managed memory for aggr…
…egations

This patch adds managed-memory-based aggregation to Spark SQL / DataFrames. Instead of working with Java objects, this new aggregation path uses `sun.misc.Unsafe` to manipulate raw memory.  This reduces the memory footprint for aggregations, resulting in fewer spills, OutOfMemoryErrors, and garbage collection pauses.  As a result, this allows for higher memory utilization.  It can also result in better cache locality since objects will be stored closer together in memory.

This feature can be eanbled by setting `spark.sql.unsafe.enabled=true`.  For now, this feature is only supported when codegen is enabled and only supports aggregations for which the grouping columns are primitive numeric types or strings and aggregated values are numeric.

### Managing memory with sun.misc.Unsafe

This patch supports both on- and off-heap managed memory.

- In on-heap mode, memory addresses are identified by the combination of a base Object and an offset within that object.
- In off-heap mode, memory is addressed directly with 64-bit long addresses.

To support both modes, functions that manipulate memory accept both `baseObject` and `baseOffset` fields.  In off-heap mode, we simply pass `null` as `baseObject`.

We allocate memory in large chunks, so memory fragmentation and allocation speed are not significant bottlenecks.

By default, we use on-heap mode.  To enable off-heap mode, set `spark.unsafe.offHeap=true`.

To track allocated memory, this patch extends `SparkEnv` with an `ExecutorMemoryManager` and supplies each `TaskContext` with a `TaskMemoryManager`.  These classes work together to track allocations and detect memory leaks.

### Compact tuple format

This patch introduces `UnsafeRow`, a compact row layout.  In this format, each tuple has three parts: a null bit set, fixed length values, and variable-length values:

![image](https://cloud.githubusercontent.com/assets/50748/7328538/2fdb65ce-ea8b-11e4-9743-6c0f02bb7d1f.png)

- Rows are always 8-byte word aligned (so their sizes will always be a multiple of 8 bytes)
- The bit set is used for null tracking:
	- Position _i_ is set if and only if field _i_ is null
	- The bit set is aligned to an 8-byte word boundary.
- Every field appears as an 8-byte word in the fixed-length values part:
	- If a field is null, we zero out the values.
	- If a field is variable-length, the word stores a relative offset (w.r.t. the base of the tuple) that points to the beginning of the field's data in the variable-length part.
- Each variable-length data type can have its own encoding:
	- For strings, the first word stores the length of the string and is followed by UTF-8 encoded bytes.  If necessary, the end of the string is padded with empty bytes in order to ensure word-alignment.

For example, a tuple that consists 3 fields of type (int, string, string), with value (null, “data”, “bricks”) would look like this:

![image](https://cloud.githubusercontent.com/assets/50748/7328526/1e21959c-ea8b-11e4-9a28-a4350fe4a7b5.png)

This format allows us to compare tuples for equality by directly comparing their raw bytes.  This also enables fast hashing of tuples.

### Hash map for performing aggregations

This patch introduces `UnsafeFixedWidthAggregationMap`, a hash map for performing aggregations where the aggregation result columns are fixed-with.  This map's keys and values are `Row` objects. `UnsafeFixedWidthAggregationMap` is implemented on top of `BytesToBytesMap`, an append-only map which supports byte-array keys and values.

`BytesToBytesMap` stores pointers to key and value tuples.  For each record with a new key, we copy the key and create the aggregation value buffer for that key and put them in a buffer. The hash table then simply stores pointers to the key and value. For each record with an existing key, we simply run the aggregation function to update the values in place.

This map is implemented using open hashing with triangular sequence probing.  Each entry stores two words in a long array: the first word stores the address of the key and the second word stores the relative offset from the key tuple to the value tuple, as well as the key's 32-bit hashcode.  By storing the full hashcode, we reduce the number of equality checks that need to be performed to handle position collisions ()since the chance of hashcode collision is much lower than position collision).

`UnsafeFixedWidthAggregationMap` allows regular Spark SQL `Row` objects to be used when probing the map.  Internally, it encodes these rows into `UnsafeRow` format using `UnsafeRowConverter`.  This conversion has a small overhead that can be eliminated in the future once we use UnsafeRows in other operators.

<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/apache/spark/5725)
<!-- Reviewable:end -->

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5725 from JoshRosen/unsafe and squashes the following commits:

eeee512 [Josh Rosen] Add converters for Null, Boolean, Byte, and Short columns.
81f34f8 [Josh Rosen] Follow 'place children last' convention for GeneratedAggregate
1bc36cc [Josh Rosen] Refactor UnsafeRowConverter to avoid unnecessary boxing.
017b2dc [Josh Rosen] Remove BytesToBytesMap.finalize()
50e9671 [Josh Rosen] Throw memory leak warning even in case of error; add warning about code duplication
70a39e4 [Josh Rosen] Split MemoryManager into ExecutorMemoryManager and TaskMemoryManager:
6e4b192 [Josh Rosen] Remove an unused method from ByteArrayMethods.
de5e001 [Josh Rosen] Fix debug vs. trace in logging message.
a19e066 [Josh Rosen] Rename unsafe Java test suites to match Scala test naming convention.
78a5b84 [Josh Rosen] Add logging to MemoryManager
ce3c565 [Josh Rosen] More comments, formatting, and code cleanup.
529e571 [Josh Rosen] Measure timeSpentResizing in nanoseconds instead of milliseconds.
3ca84b2 [Josh Rosen] Only zero the used portion of groupingKeyConversionScratchSpace
162caf7 [Josh Rosen] Fix test compilation
b45f070 [Josh Rosen] Don't redundantly store the offset from key to value, since we can compute this from the key size.
a8e4a3f [Josh Rosen] Introduce MemoryManager interface; add to SparkEnv.
0925847 [Josh Rosen] Disable MiMa checks for new unsafe module
cde4132 [Josh Rosen] Add missing pom.xml
9c19fc0 [Josh Rosen] Add configuration options for heap vs. offheap
6ffdaa1 [Josh Rosen] Null handling improvements in UnsafeRow.
31eaabc [Josh Rosen] Lots of TODO and doc cleanup.
a95291e [Josh Rosen] Cleanups to string handling code
afe8dca [Josh Rosen] Some Javadoc cleanup
f3dcbfe [Josh Rosen] More mod replacement
854201a [Josh Rosen] Import and comment cleanup
06e929d [Josh Rosen] More warning cleanup
ef6b3d3 [Josh Rosen] Fix a bunch of FindBugs and IntelliJ inspections
29a7575 [Josh Rosen] Remove debug logging
49aed30 [Josh Rosen] More long -> int conversion.
b26f1d3 [Josh Rosen] Fix bug in murmur hash implementation.
765243d [Josh Rosen] Enable optional performance metrics for hash map.
23a440a [Josh Rosen] Bump up default hash map size
628f936 [Josh Rosen] Use ints intead of longs for indexing.
92d5a06 [Josh Rosen] Address a number of minor code review comments.
1f4b716 [Josh Rosen] Merge Unsafe code into the regular GeneratedAggregate, guarded by a configuration flag; integrate planner support and re-enable all tests.
d85eeff [Josh Rosen] Add basic sanity test for UnsafeFixedWidthAggregationMap
bade966 [Josh Rosen] Comment update (bumping to refresh GitHub cache...)
b3eaccd [Josh Rosen] Extract aggregation map into its own class.
d2bb986 [Josh Rosen] Update to implement new Row methods added upstream
58ac393 [Josh Rosen] Use UNSAFE allocator in GeneratedAggregate (TODO: make this configurable)
7df6008 [Josh Rosen] Optimizations related to zeroing out memory:
c1b3813 [Josh Rosen] Fix bug in UnsafeMemoryAllocator.free():
738fa33 [Josh Rosen] Add feature flag to guard UnsafeGeneratedAggregate
c55bf66 [Josh Rosen] Free buffer once iterator has been fully consumed.
62ab054 [Josh Rosen] Optimize for fact that get() is only called on String columns.
c7f0b56 [Josh Rosen] Reuse UnsafeRow pointer in UnsafeRowConverter
ae39694 [Josh Rosen] Add finalizer as "cleanup method of last resort"
c754ae1 [Josh Rosen] Now that the store*() contract has been stregthened, we can remove an extra lookup
f764d13 [Josh Rosen] Simplify address + length calculation in Location.
079f1bf [Josh Rosen] Some clarification of the BytesToBytesMap.lookup() / set() contract.
1a483c5 [Josh Rosen] First version that passes some aggregation tests:
fc4c3a8 [Josh Rosen] Sketch how the converters will be used in UnsafeGeneratedAggregate
53ba9b7 [Josh Rosen] Start prototyping Java Row -> UnsafeRow converters
1ff814d [Josh Rosen] Add reminder to free memory on iterator completion
8a8f9df [Josh Rosen] Add skeleton for GeneratedAggregate integration.
5d55cef [Josh Rosen] Add skeleton for Row implementation.
f03e9c1 [Josh Rosen] Play around with Unsafe implementations of more string methods.
ab68e08 [Josh Rosen] Begin merging the UTF8String implementations.
480a74a [Josh Rosen] Initial import of code from Databricks unsafe utils repo.

jeanlyn added a commit to jeanlyn/spark that referenced this pull request Jun 12, 2015

[SPARK-7288] Suppress compiler warnings due to use of sun.misc.Unsafe…
…; add facade in front of Unsafe; remove use of Unsafe.setMemory

This patch suppresses compiler warnings due to our use of `sun.misc.Unsafe` (introduced in #5725).  These warnings can only be suppressed via the `-XDignore.symbol.file` javac flag; the `SuppressWarnings` annotation won't work for these.

In order to restrict uses of this compiler flag to the `unsafe` module, I placed a facade in front of `Unsafe` so that other modules won't call it directly. This facade also will also help us to avoid accidental usage of deprecated Unsafe methods or methods that aren't supported in Java 6.

I also removed an unnecessary use of `Unsafe.setMemory`, which isn't present in certain versions of Java 6, and excluded the new `unsafe` module from Javadoc.

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5814 from JoshRosen/unsafe-compiler-warnings-fixes and squashes the following commits:

9e8c483 [Josh Rosen] Exclude new unsafe module from Javadoc
ba75ecf [Josh Rosen] Only apply -XDignore.symbol.file flag in unsafe project.
7403345 [Josh Rosen] Put facade in front of Unsafe.
50230c0 [Josh Rosen] Remove usage of Unsafe.setMemory
96d41c9 [Josh Rosen] Use -XDignore.symbol.file to suppress warnings about sun.misc.Unsafe usage

jeanlyn added a commit to jeanlyn/spark that referenced this pull request Jun 12, 2015

[SPARK-7081] Faster sort-based shuffle path using binary processing c…
…ache-aware sort

This patch introduces a new shuffle manager that enhances the existing sort-based shuffle with a new cache-friendly sort algorithm that operates directly on binary data. The goals of this patch are to lower memory usage and Java object overheads during shuffle and to speed up sorting. It also lays groundwork for follow-up patches that will enable end-to-end processing of serialized records.

The new shuffle manager, `UnsafeShuffleManager`, can be enabled by setting `spark.shuffle.manager=tungsten-sort` in SparkConf.

The new shuffle manager uses directly-managed memory to implement several performance optimizations for certain types of shuffles. In cases where the new performance optimizations cannot be applied, the new shuffle manager delegates to SortShuffleManager to handle those shuffles.

UnsafeShuffleManager's optimizations will apply when _all_ of the following conditions hold:

 - The shuffle dependency specifies no aggregation or output ordering.
 - The shuffle serializer supports relocation of serialized values (this is currently supported
   by KryoSerializer and Spark SQL's custom serializers).
 - The shuffle produces fewer than 16777216 output partitions.
 - No individual record is larger than 128 MB when serialized.

In addition, extra spill-merging optimizations are automatically applied when the shuffle compression codec supports concatenation of serialized streams. This is currently supported by Spark's LZF serializer.

At a high-level, UnsafeShuffleManager's design is similar to Spark's existing SortShuffleManager.  In sort-based shuffle, incoming records are sorted according to their target partition ids, then written to a single map output file. Reducers fetch contiguous regions of this file in order to read their portion of the map output. In cases where the map output data is too large to fit in memory, sorted subsets of the output can are spilled to disk and those on-disk files are merged to produce the final output file.

UnsafeShuffleManager optimizes this process in several ways:

 - Its sort operates on serialized binary data rather than Java objects, which reduces memory consumption and GC overheads. This optimization requires the record serializer to have certain properties to allow serialized records to be re-ordered without requiring deserialization.  See SPARK-4550, where this optimization was first proposed and implemented, for more details.

 - It uses a specialized cache-efficient sorter (UnsafeShuffleExternalSorter) that sorts arrays of compressed record pointers and partition ids. By using only 8 bytes of space per record in the sorting array, this fits more of the array into cache.

 - The spill merging procedure operates on blocks of serialized records that belong to the same partition and does not need to deserialize records during the merge.

 - When the spill compression codec supports concatenation of compressed data, the spill merge simply concatenates the serialized and compressed spill partitions to produce the final output partition.  This allows efficient data copying methods, like NIO's `transferTo`, to be used and avoids the need to allocate decompression or copying buffers during the merge.

The shuffle read path is unchanged.

This patch is similar to [SPARK-4550](http://issues.apache.org/jira/browse/SPARK-4550) / #4450 but uses a slightly different implementation. The `unsafe`-based implementation featured in this patch lays the groundwork for followup patches that will enable sorting to operate on serialized data pages that will be prepared by Spark SQL's new `unsafe` operators (such as the new aggregation operator introduced in #5725).

### Future work

There are several tasks that build upon this patch, which will be left to future work:

- [SPARK-7271](https://issues.apache.org/jira/browse/SPARK-7271) Redesign / extend the shuffle interfaces to accept binary data as input. The goal here is to let us bypass serialization steps in cases where the sort input is produced by an operator that operates directly on binary data.
- Extension / redesign of the `Serializer` API. We can add new methods which allow serializers to determine the size requirements for serializing objects and for serializing objects directly to a specified memory address (similar to how `UnsafeRowConverter` works in Spark SQL).

<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/apache/spark/5868)
<!-- Reviewable:end -->

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5868 from JoshRosen/unsafe-sort and squashes the following commits:

ef0a86e [Josh Rosen] Fix scalastyle errors
7610f2f [Josh Rosen] Add tests for proper cleanup of shuffle data.
d494ffe [Josh Rosen] Fix deserialization of JavaSerializer instances.
52a9981 [Josh Rosen] Fix some bugs in the address packing code.
51812a7 [Josh Rosen] Change shuffle manager sort name to tungsten-sort
4023fa4 [Josh Rosen] Add @Private annotation to some Java classes.
de40b9d [Josh Rosen] More comments to try to explain metrics code
df07699 [Josh Rosen] Attempt to clarify confusing metrics update code
5e189c6 [Josh Rosen] Track time spend closing / flushing files; split TimeTrackingOutputStream into separate file.
d5779c6 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
c2ce78e [Josh Rosen] Fix a missed usage of MAX_PARTITION_ID
e3b8855 [Josh Rosen] Cleanup in UnsafeShuffleWriter
4a2c785 [Josh Rosen] rename 'sort buffer' to 'pointer array'
6276168 [Josh Rosen] Remove ability to disable spilling in UnsafeShuffleExternalSorter.
57312c9 [Josh Rosen] Clarify fileBufferSize units
2d4e4f4 [Josh Rosen] Address some minor comments in UnsafeShuffleExternalSorter.
fdcac08 [Josh Rosen] Guard against overflow when expanding sort buffer.
85da63f [Josh Rosen] Cleanup in UnsafeShuffleSorterIterator.
0ad34da [Josh Rosen] Fix off-by-one in nextInt() call
56781a1 [Josh Rosen] Rename UnsafeShuffleSorter to UnsafeShuffleInMemorySorter
e995d1a [Josh Rosen] Introduce MAX_SHUFFLE_OUTPUT_PARTITIONS.
e58a6b4 [Josh Rosen] Add more tests for PackedRecordPointer encoding.
4f0b770 [Josh Rosen] Attempt to implement proper shuffle write metrics.
d4e6d89 [Josh Rosen] Update to bit shifting constants
69d5899 [Josh Rosen] Remove some unnecessary override vals
8531286 [Josh Rosen] Add tests that automatically trigger spills.
7c953f9 [Josh Rosen] Add test that covers UnsafeShuffleSortDataFormat.swap().
e1855e5 [Josh Rosen] Fix a handful of misc. IntelliJ inspections
39434f9 [Josh Rosen] Avoid integer multiplication overflow in getMemoryUsage (thanks FindBugs!)
1e3ad52 [Josh Rosen] Delete unused ByteBufferOutputStream class.
ea4f85f [Josh Rosen] Roll back an unnecessary change in Spillable.
ae538dc [Josh Rosen] Document UnsafeShuffleManager.
ec6d626 [Josh Rosen] Add notes on maximum # of supported shuffle partitions.
0d4d199 [Josh Rosen] Bump up shuffle.memoryFraction to make tests pass.
b3b1924 [Josh Rosen] Properly implement close() and flush() in DummySerializerInstance.
1ef56c7 [Josh Rosen] Revise compression codec support in merger; test cross product of configurations.
b57c17f [Josh Rosen] Disable some overly-verbose logs that rendered DEBUG useless.
f780fb1 [Josh Rosen] Add test demonstrating which compression codecs support concatenation.
4a01c45 [Josh Rosen] Remove unnecessary log message
27b18b0 [Josh Rosen] That for inserting records AT the max record size.
fcd9a3c [Josh Rosen] Add notes + tests for maximum record / page sizes.
9d1ee7c [Josh Rosen] Fix MiMa excludes for ShuffleWriter change
fd4bb9e [Josh Rosen] Use own ByteBufferOutputStream rather than Kryo's
67d25ba [Josh Rosen] Update Exchange operator's copying logic to account for new shuffle manager
8f5061a [Josh Rosen] Strengthen assertion to check partitioning
01afc74 [Josh Rosen] Actually read data in UnsafeShuffleWriterSuite
1929a74 [Josh Rosen] Update to reflect upstream ShuffleBlockManager -> ShuffleBlockResolver rename.
e8718dd [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
9b7ebed [Josh Rosen] More defensive programming RE: cleaning up spill files and memory after errors
7cd013b [Josh Rosen] Begin refactoring to enable proper tests for spilling.
722849b [Josh Rosen] Add workaround for transferTo() bug in merging code; refactor tests.
9883e30 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
b95e642 [Josh Rosen] Refactor and document logic that decides when to spill.
1ce1300 [Josh Rosen] More minor cleanup
5e8cf75 [Josh Rosen] More minor cleanup
e67f1ea [Josh Rosen] Remove upper type bound in ShuffleWriter interface.
cfe0ec4 [Josh Rosen] Address a number of minor review comments:
8a6fe52 [Josh Rosen] Rename UnsafeShuffleSpillWriter to UnsafeShuffleExternalSorter
11feeb6 [Josh Rosen] Update TODOs related to shuffle write metrics.
b674412 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
aaea17b [Josh Rosen] Add comments to UnsafeShuffleSpillWriter.
4f70141 [Josh Rosen] Fix merging; now passes UnsafeShuffleSuite tests.
133c8c9 [Josh Rosen] WIP towards testing UnsafeShuffleWriter.
f480fb2 [Josh Rosen] WIP in mega-refactoring towards shuffle-specific sort.
57f1ec0 [Josh Rosen] WIP towards packed record pointers for use in optimized shuffle sort.
69232fd [Josh Rosen] Enable compressible address encoding for off-heap mode.
7ee918e [Josh Rosen] Re-order imports in tests
3aeaff7 [Josh Rosen] More refactoring and cleanup; begin cleaning iterator interfaces
3490512 [Josh Rosen] Misc. cleanup
f156a8f [Josh Rosen] Hacky metrics integration; refactor some interfaces.
2776aca [Josh Rosen] First passing test for ExternalSorter.
5e100b2 [Josh Rosen] Super-messy WIP on external sort
595923a [Josh Rosen] Remove some unused variables.
8958584 [Josh Rosen] Fix bug in calculating free space in current page.
f17fa8f [Josh Rosen] Add missing newline
c2fca17 [Josh Rosen] Small refactoring of SerializerPropertiesSuite to enable test re-use:
b8a09fe [Josh Rosen] Back out accidental log4j.properties change
bfc12d3 [Josh Rosen] Add tests for serializer relocation property.
240864c [Josh Rosen] Remove PrefixComputer and require prefix to be specified as part of insert()
1433b42 [Josh Rosen] Store record length as int instead of long.
026b497 [Josh Rosen] Re-use a buffer in UnsafeShuffleWriter
0748458 [Josh Rosen] Port UnsafeShuffleWriter to Java.
87e721b [Josh Rosen] Renaming and comments
d3cc310 [Josh Rosen] Flag that SparkSqlSerializer2 supports relocation
e2d96ca [Josh Rosen] Expand serializer API and use new function to help control when new UnsafeShuffle path is used.
e267cee [Josh Rosen] Fix compilation of UnsafeSorterSuite
9c6cf58 [Josh Rosen] Refactor to use DiskBlockObjectWriter.
253f13e [Josh Rosen] More cleanup
8e3ec20 [Josh Rosen] Begin code cleanup.
4d2f5e1 [Josh Rosen] WIP
3db12de [Josh Rosen] Minor simplification and sanity checks in UnsafeSorter
767d3ca [Josh Rosen] Fix invalid range in UnsafeSorter.
e900152 [Josh Rosen] Add test for empty iterator in UnsafeSorter
57a4ea0 [Josh Rosen] Make initialSize configurable in UnsafeSorter
abf7bfe [Josh Rosen] Add basic test case.
81d52c5 [Josh Rosen] WIP on UnsafeSorter

nemccarthy added a commit to nemccarthy/spark that referenced this pull request Jun 19, 2015

[SPARK-7076][SPARK-7077][SPARK-7080][SQL] Use managed memory for aggr…
…egations

This patch adds managed-memory-based aggregation to Spark SQL / DataFrames. Instead of working with Java objects, this new aggregation path uses `sun.misc.Unsafe` to manipulate raw memory.  This reduces the memory footprint for aggregations, resulting in fewer spills, OutOfMemoryErrors, and garbage collection pauses.  As a result, this allows for higher memory utilization.  It can also result in better cache locality since objects will be stored closer together in memory.

This feature can be eanbled by setting `spark.sql.unsafe.enabled=true`.  For now, this feature is only supported when codegen is enabled and only supports aggregations for which the grouping columns are primitive numeric types or strings and aggregated values are numeric.

### Managing memory with sun.misc.Unsafe

This patch supports both on- and off-heap managed memory.

- In on-heap mode, memory addresses are identified by the combination of a base Object and an offset within that object.
- In off-heap mode, memory is addressed directly with 64-bit long addresses.

To support both modes, functions that manipulate memory accept both `baseObject` and `baseOffset` fields.  In off-heap mode, we simply pass `null` as `baseObject`.

We allocate memory in large chunks, so memory fragmentation and allocation speed are not significant bottlenecks.

By default, we use on-heap mode.  To enable off-heap mode, set `spark.unsafe.offHeap=true`.

To track allocated memory, this patch extends `SparkEnv` with an `ExecutorMemoryManager` and supplies each `TaskContext` with a `TaskMemoryManager`.  These classes work together to track allocations and detect memory leaks.

### Compact tuple format

This patch introduces `UnsafeRow`, a compact row layout.  In this format, each tuple has three parts: a null bit set, fixed length values, and variable-length values:

![image](https://cloud.githubusercontent.com/assets/50748/7328538/2fdb65ce-ea8b-11e4-9743-6c0f02bb7d1f.png)

- Rows are always 8-byte word aligned (so their sizes will always be a multiple of 8 bytes)
- The bit set is used for null tracking:
	- Position _i_ is set if and only if field _i_ is null
	- The bit set is aligned to an 8-byte word boundary.
- Every field appears as an 8-byte word in the fixed-length values part:
	- If a field is null, we zero out the values.
	- If a field is variable-length, the word stores a relative offset (w.r.t. the base of the tuple) that points to the beginning of the field's data in the variable-length part.
- Each variable-length data type can have its own encoding:
	- For strings, the first word stores the length of the string and is followed by UTF-8 encoded bytes.  If necessary, the end of the string is padded with empty bytes in order to ensure word-alignment.

For example, a tuple that consists 3 fields of type (int, string, string), with value (null, “data”, “bricks”) would look like this:

![image](https://cloud.githubusercontent.com/assets/50748/7328526/1e21959c-ea8b-11e4-9a28-a4350fe4a7b5.png)

This format allows us to compare tuples for equality by directly comparing their raw bytes.  This also enables fast hashing of tuples.

### Hash map for performing aggregations

This patch introduces `UnsafeFixedWidthAggregationMap`, a hash map for performing aggregations where the aggregation result columns are fixed-with.  This map's keys and values are `Row` objects. `UnsafeFixedWidthAggregationMap` is implemented on top of `BytesToBytesMap`, an append-only map which supports byte-array keys and values.

`BytesToBytesMap` stores pointers to key and value tuples.  For each record with a new key, we copy the key and create the aggregation value buffer for that key and put them in a buffer. The hash table then simply stores pointers to the key and value. For each record with an existing key, we simply run the aggregation function to update the values in place.

This map is implemented using open hashing with triangular sequence probing.  Each entry stores two words in a long array: the first word stores the address of the key and the second word stores the relative offset from the key tuple to the value tuple, as well as the key's 32-bit hashcode.  By storing the full hashcode, we reduce the number of equality checks that need to be performed to handle position collisions ()since the chance of hashcode collision is much lower than position collision).

`UnsafeFixedWidthAggregationMap` allows regular Spark SQL `Row` objects to be used when probing the map.  Internally, it encodes these rows into `UnsafeRow` format using `UnsafeRowConverter`.  This conversion has a small overhead that can be eliminated in the future once we use UnsafeRows in other operators.

<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/apache/spark/5725)
<!-- Reviewable:end -->

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5725 from JoshRosen/unsafe and squashes the following commits:

eeee512 [Josh Rosen] Add converters for Null, Boolean, Byte, and Short columns.
81f34f8 [Josh Rosen] Follow 'place children last' convention for GeneratedAggregate
1bc36cc [Josh Rosen] Refactor UnsafeRowConverter to avoid unnecessary boxing.
017b2dc [Josh Rosen] Remove BytesToBytesMap.finalize()
50e9671 [Josh Rosen] Throw memory leak warning even in case of error; add warning about code duplication
70a39e4 [Josh Rosen] Split MemoryManager into ExecutorMemoryManager and TaskMemoryManager:
6e4b192 [Josh Rosen] Remove an unused method from ByteArrayMethods.
de5e001 [Josh Rosen] Fix debug vs. trace in logging message.
a19e066 [Josh Rosen] Rename unsafe Java test suites to match Scala test naming convention.
78a5b84 [Josh Rosen] Add logging to MemoryManager
ce3c565 [Josh Rosen] More comments, formatting, and code cleanup.
529e571 [Josh Rosen] Measure timeSpentResizing in nanoseconds instead of milliseconds.
3ca84b2 [Josh Rosen] Only zero the used portion of groupingKeyConversionScratchSpace
162caf7 [Josh Rosen] Fix test compilation
b45f070 [Josh Rosen] Don't redundantly store the offset from key to value, since we can compute this from the key size.
a8e4a3f [Josh Rosen] Introduce MemoryManager interface; add to SparkEnv.
0925847 [Josh Rosen] Disable MiMa checks for new unsafe module
cde4132 [Josh Rosen] Add missing pom.xml
9c19fc0 [Josh Rosen] Add configuration options for heap vs. offheap
6ffdaa1 [Josh Rosen] Null handling improvements in UnsafeRow.
31eaabc [Josh Rosen] Lots of TODO and doc cleanup.
a95291e [Josh Rosen] Cleanups to string handling code
afe8dca [Josh Rosen] Some Javadoc cleanup
f3dcbfe [Josh Rosen] More mod replacement
854201a [Josh Rosen] Import and comment cleanup
06e929d [Josh Rosen] More warning cleanup
ef6b3d3 [Josh Rosen] Fix a bunch of FindBugs and IntelliJ inspections
29a7575 [Josh Rosen] Remove debug logging
49aed30 [Josh Rosen] More long -> int conversion.
b26f1d3 [Josh Rosen] Fix bug in murmur hash implementation.
765243d [Josh Rosen] Enable optional performance metrics for hash map.
23a440a [Josh Rosen] Bump up default hash map size
628f936 [Josh Rosen] Use ints intead of longs for indexing.
92d5a06 [Josh Rosen] Address a number of minor code review comments.
1f4b716 [Josh Rosen] Merge Unsafe code into the regular GeneratedAggregate, guarded by a configuration flag; integrate planner support and re-enable all tests.
d85eeff [Josh Rosen] Add basic sanity test for UnsafeFixedWidthAggregationMap
bade966 [Josh Rosen] Comment update (bumping to refresh GitHub cache...)
b3eaccd [Josh Rosen] Extract aggregation map into its own class.
d2bb986 [Josh Rosen] Update to implement new Row methods added upstream
58ac393 [Josh Rosen] Use UNSAFE allocator in GeneratedAggregate (TODO: make this configurable)
7df6008 [Josh Rosen] Optimizations related to zeroing out memory:
c1b3813 [Josh Rosen] Fix bug in UnsafeMemoryAllocator.free():
738fa33 [Josh Rosen] Add feature flag to guard UnsafeGeneratedAggregate
c55bf66 [Josh Rosen] Free buffer once iterator has been fully consumed.
62ab054 [Josh Rosen] Optimize for fact that get() is only called on String columns.
c7f0b56 [Josh Rosen] Reuse UnsafeRow pointer in UnsafeRowConverter
ae39694 [Josh Rosen] Add finalizer as "cleanup method of last resort"
c754ae1 [Josh Rosen] Now that the store*() contract has been stregthened, we can remove an extra lookup
f764d13 [Josh Rosen] Simplify address + length calculation in Location.
079f1bf [Josh Rosen] Some clarification of the BytesToBytesMap.lookup() / set() contract.
1a483c5 [Josh Rosen] First version that passes some aggregation tests:
fc4c3a8 [Josh Rosen] Sketch how the converters will be used in UnsafeGeneratedAggregate
53ba9b7 [Josh Rosen] Start prototyping Java Row -> UnsafeRow converters
1ff814d [Josh Rosen] Add reminder to free memory on iterator completion
8a8f9df [Josh Rosen] Add skeleton for GeneratedAggregate integration.
5d55cef [Josh Rosen] Add skeleton for Row implementation.
f03e9c1 [Josh Rosen] Play around with Unsafe implementations of more string methods.
ab68e08 [Josh Rosen] Begin merging the UTF8String implementations.
480a74a [Josh Rosen] Initial import of code from Databricks unsafe utils repo.

nemccarthy added a commit to nemccarthy/spark that referenced this pull request Jun 19, 2015

[SPARK-7288] Suppress compiler warnings due to use of sun.misc.Unsafe…
…; add facade in front of Unsafe; remove use of Unsafe.setMemory

This patch suppresses compiler warnings due to our use of `sun.misc.Unsafe` (introduced in #5725).  These warnings can only be suppressed via the `-XDignore.symbol.file` javac flag; the `SuppressWarnings` annotation won't work for these.

In order to restrict uses of this compiler flag to the `unsafe` module, I placed a facade in front of `Unsafe` so that other modules won't call it directly. This facade also will also help us to avoid accidental usage of deprecated Unsafe methods or methods that aren't supported in Java 6.

I also removed an unnecessary use of `Unsafe.setMemory`, which isn't present in certain versions of Java 6, and excluded the new `unsafe` module from Javadoc.

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5814 from JoshRosen/unsafe-compiler-warnings-fixes and squashes the following commits:

9e8c483 [Josh Rosen] Exclude new unsafe module from Javadoc
ba75ecf [Josh Rosen] Only apply -XDignore.symbol.file flag in unsafe project.
7403345 [Josh Rosen] Put facade in front of Unsafe.
50230c0 [Josh Rosen] Remove usage of Unsafe.setMemory
96d41c9 [Josh Rosen] Use -XDignore.symbol.file to suppress warnings about sun.misc.Unsafe usage

nemccarthy added a commit to nemccarthy/spark that referenced this pull request Jun 19, 2015

[SPARK-7081] Faster sort-based shuffle path using binary processing c…
…ache-aware sort

This patch introduces a new shuffle manager that enhances the existing sort-based shuffle with a new cache-friendly sort algorithm that operates directly on binary data. The goals of this patch are to lower memory usage and Java object overheads during shuffle and to speed up sorting. It also lays groundwork for follow-up patches that will enable end-to-end processing of serialized records.

The new shuffle manager, `UnsafeShuffleManager`, can be enabled by setting `spark.shuffle.manager=tungsten-sort` in SparkConf.

The new shuffle manager uses directly-managed memory to implement several performance optimizations for certain types of shuffles. In cases where the new performance optimizations cannot be applied, the new shuffle manager delegates to SortShuffleManager to handle those shuffles.

UnsafeShuffleManager's optimizations will apply when _all_ of the following conditions hold:

 - The shuffle dependency specifies no aggregation or output ordering.
 - The shuffle serializer supports relocation of serialized values (this is currently supported
   by KryoSerializer and Spark SQL's custom serializers).
 - The shuffle produces fewer than 16777216 output partitions.
 - No individual record is larger than 128 MB when serialized.

In addition, extra spill-merging optimizations are automatically applied when the shuffle compression codec supports concatenation of serialized streams. This is currently supported by Spark's LZF serializer.

At a high-level, UnsafeShuffleManager's design is similar to Spark's existing SortShuffleManager.  In sort-based shuffle, incoming records are sorted according to their target partition ids, then written to a single map output file. Reducers fetch contiguous regions of this file in order to read their portion of the map output. In cases where the map output data is too large to fit in memory, sorted subsets of the output can are spilled to disk and those on-disk files are merged to produce the final output file.

UnsafeShuffleManager optimizes this process in several ways:

 - Its sort operates on serialized binary data rather than Java objects, which reduces memory consumption and GC overheads. This optimization requires the record serializer to have certain properties to allow serialized records to be re-ordered without requiring deserialization.  See SPARK-4550, where this optimization was first proposed and implemented, for more details.

 - It uses a specialized cache-efficient sorter (UnsafeShuffleExternalSorter) that sorts arrays of compressed record pointers and partition ids. By using only 8 bytes of space per record in the sorting array, this fits more of the array into cache.

 - The spill merging procedure operates on blocks of serialized records that belong to the same partition and does not need to deserialize records during the merge.

 - When the spill compression codec supports concatenation of compressed data, the spill merge simply concatenates the serialized and compressed spill partitions to produce the final output partition.  This allows efficient data copying methods, like NIO's `transferTo`, to be used and avoids the need to allocate decompression or copying buffers during the merge.

The shuffle read path is unchanged.

This patch is similar to [SPARK-4550](http://issues.apache.org/jira/browse/SPARK-4550) / #4450 but uses a slightly different implementation. The `unsafe`-based implementation featured in this patch lays the groundwork for followup patches that will enable sorting to operate on serialized data pages that will be prepared by Spark SQL's new `unsafe` operators (such as the new aggregation operator introduced in #5725).

### Future work

There are several tasks that build upon this patch, which will be left to future work:

- [SPARK-7271](https://issues.apache.org/jira/browse/SPARK-7271) Redesign / extend the shuffle interfaces to accept binary data as input. The goal here is to let us bypass serialization steps in cases where the sort input is produced by an operator that operates directly on binary data.
- Extension / redesign of the `Serializer` API. We can add new methods which allow serializers to determine the size requirements for serializing objects and for serializing objects directly to a specified memory address (similar to how `UnsafeRowConverter` works in Spark SQL).

<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/apache/spark/5868)
<!-- Reviewable:end -->

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5868 from JoshRosen/unsafe-sort and squashes the following commits:

ef0a86e [Josh Rosen] Fix scalastyle errors
7610f2f [Josh Rosen] Add tests for proper cleanup of shuffle data.
d494ffe [Josh Rosen] Fix deserialization of JavaSerializer instances.
52a9981 [Josh Rosen] Fix some bugs in the address packing code.
51812a7 [Josh Rosen] Change shuffle manager sort name to tungsten-sort
4023fa4 [Josh Rosen] Add @Private annotation to some Java classes.
de40b9d [Josh Rosen] More comments to try to explain metrics code
df07699 [Josh Rosen] Attempt to clarify confusing metrics update code
5e189c6 [Josh Rosen] Track time spend closing / flushing files; split TimeTrackingOutputStream into separate file.
d5779c6 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
c2ce78e [Josh Rosen] Fix a missed usage of MAX_PARTITION_ID
e3b8855 [Josh Rosen] Cleanup in UnsafeShuffleWriter
4a2c785 [Josh Rosen] rename 'sort buffer' to 'pointer array'
6276168 [Josh Rosen] Remove ability to disable spilling in UnsafeShuffleExternalSorter.
57312c9 [Josh Rosen] Clarify fileBufferSize units
2d4e4f4 [Josh Rosen] Address some minor comments in UnsafeShuffleExternalSorter.
fdcac08 [Josh Rosen] Guard against overflow when expanding sort buffer.
85da63f [Josh Rosen] Cleanup in UnsafeShuffleSorterIterator.
0ad34da [Josh Rosen] Fix off-by-one in nextInt() call
56781a1 [Josh Rosen] Rename UnsafeShuffleSorter to UnsafeShuffleInMemorySorter
e995d1a [Josh Rosen] Introduce MAX_SHUFFLE_OUTPUT_PARTITIONS.
e58a6b4 [Josh Rosen] Add more tests for PackedRecordPointer encoding.
4f0b770 [Josh Rosen] Attempt to implement proper shuffle write metrics.
d4e6d89 [Josh Rosen] Update to bit shifting constants
69d5899 [Josh Rosen] Remove some unnecessary override vals
8531286 [Josh Rosen] Add tests that automatically trigger spills.
7c953f9 [Josh Rosen] Add test that covers UnsafeShuffleSortDataFormat.swap().
e1855e5 [Josh Rosen] Fix a handful of misc. IntelliJ inspections
39434f9 [Josh Rosen] Avoid integer multiplication overflow in getMemoryUsage (thanks FindBugs!)
1e3ad52 [Josh Rosen] Delete unused ByteBufferOutputStream class.
ea4f85f [Josh Rosen] Roll back an unnecessary change in Spillable.
ae538dc [Josh Rosen] Document UnsafeShuffleManager.
ec6d626 [Josh Rosen] Add notes on maximum # of supported shuffle partitions.
0d4d199 [Josh Rosen] Bump up shuffle.memoryFraction to make tests pass.
b3b1924 [Josh Rosen] Properly implement close() and flush() in DummySerializerInstance.
1ef56c7 [Josh Rosen] Revise compression codec support in merger; test cross product of configurations.
b57c17f [Josh Rosen] Disable some overly-verbose logs that rendered DEBUG useless.
f780fb1 [Josh Rosen] Add test demonstrating which compression codecs support concatenation.
4a01c45 [Josh Rosen] Remove unnecessary log message
27b18b0 [Josh Rosen] That for inserting records AT the max record size.
fcd9a3c [Josh Rosen] Add notes + tests for maximum record / page sizes.
9d1ee7c [Josh Rosen] Fix MiMa excludes for ShuffleWriter change
fd4bb9e [Josh Rosen] Use own ByteBufferOutputStream rather than Kryo's
67d25ba [Josh Rosen] Update Exchange operator's copying logic to account for new shuffle manager
8f5061a [Josh Rosen] Strengthen assertion to check partitioning
01afc74 [Josh Rosen] Actually read data in UnsafeShuffleWriterSuite
1929a74 [Josh Rosen] Update to reflect upstream ShuffleBlockManager -> ShuffleBlockResolver rename.
e8718dd [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
9b7ebed [Josh Rosen] More defensive programming RE: cleaning up spill files and memory after errors
7cd013b [Josh Rosen] Begin refactoring to enable proper tests for spilling.
722849b [Josh Rosen] Add workaround for transferTo() bug in merging code; refactor tests.
9883e30 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
b95e642 [Josh Rosen] Refactor and document logic that decides when to spill.
1ce1300 [Josh Rosen] More minor cleanup
5e8cf75 [Josh Rosen] More minor cleanup
e67f1ea [Josh Rosen] Remove upper type bound in ShuffleWriter interface.
cfe0ec4 [Josh Rosen] Address a number of minor review comments:
8a6fe52 [Josh Rosen] Rename UnsafeShuffleSpillWriter to UnsafeShuffleExternalSorter
11feeb6 [Josh Rosen] Update TODOs related to shuffle write metrics.
b674412 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-sort
aaea17b [Josh Rosen] Add comments to UnsafeShuffleSpillWriter.
4f70141 [Josh Rosen] Fix merging; now passes UnsafeShuffleSuite tests.
133c8c9 [Josh Rosen] WIP towards testing UnsafeShuffleWriter.
f480fb2 [Josh Rosen] WIP in mega-refactoring towards shuffle-specific sort.
57f1ec0 [Josh Rosen] WIP towards packed record pointers for use in optimized shuffle sort.
69232fd [Josh Rosen] Enable compressible address encoding for off-heap mode.
7ee918e [Josh Rosen] Re-order imports in tests
3aeaff7 [Josh Rosen] More refactoring and cleanup; begin cleaning iterator interfaces
3490512 [Josh Rosen] Misc. cleanup
f156a8f [Josh Rosen] Hacky metrics integration; refactor some interfaces.
2776aca [Josh Rosen] First passing test for ExternalSorter.
5e100b2 [Josh Rosen] Super-messy WIP on external sort
595923a [Josh Rosen] Remove some unused variables.
8958584 [Josh Rosen] Fix bug in calculating free space in current page.
f17fa8f [Josh Rosen] Add missing newline
c2fca17 [Josh Rosen] Small refactoring of SerializerPropertiesSuite to enable test re-use:
b8a09fe [Josh Rosen] Back out accidental log4j.properties change
bfc12d3 [Josh Rosen] Add tests for serializer relocation property.
240864c [Josh Rosen] Remove PrefixComputer and require prefix to be specified as part of insert()
1433b42 [Josh Rosen] Store record length as int instead of long.
026b497 [Josh Rosen] Re-use a buffer in UnsafeShuffleWriter
0748458 [Josh Rosen] Port UnsafeShuffleWriter to Java.
87e721b [Josh Rosen] Renaming and comments
d3cc310 [Josh Rosen] Flag that SparkSqlSerializer2 supports relocation
e2d96ca [Josh Rosen] Expand serializer API and use new function to help control when new UnsafeShuffle path is used.
e267cee [Josh Rosen] Fix compilation of UnsafeSorterSuite
9c6cf58 [Josh Rosen] Refactor to use DiskBlockObjectWriter.
253f13e [Josh Rosen] More cleanup
8e3ec20 [Josh Rosen] Begin code cleanup.
4d2f5e1 [Josh Rosen] WIP
3db12de [Josh Rosen] Minor simplification and sanity checks in UnsafeSorter
767d3ca [Josh Rosen] Fix invalid range in UnsafeSorter.
e900152 [Josh Rosen] Add test for empty iterator in UnsafeSorter
57a4ea0 [Josh Rosen] Make initialSize configurable in UnsafeSorter
abf7bfe [Josh Rosen] Add basic test case.
81d52c5 [Josh Rosen] WIP on UnsafeSorter
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment