[SPARK-37037][SQL] Improve byte array sort by unify compareTo function of UTF8String and ByteArray #34310
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Kubernetes integration test starting |
|
Kubernetes integration test status failure |
|
Test build #144352 has finished for PR 34310 at commit
|
|
Kubernetes integration test starting |
|
Kubernetes integration test status failure |
|
Test build #144362 has finished for PR 34310 at commit
|
Contributor
Author
|
retest this please |
Contributor
Author
|
cc @xkrogen @srowen @kiszk @cloud-fan @JoshRosen if you have time to take another look |
|
Kubernetes integration test starting |
|
Kubernetes integration test status failure |
|
Test build #144367 has finished for PR 34310 at commit
|
srowen
reviewed
Oct 18, 2021
| } | ||
|
|
||
| public static int compareBinary(byte[] leftBase, byte[] rightBase) { | ||
| return compareBinary(leftBase, Platform.BYTE_ARRAY_OFFSET, leftBase.length, |
Member
There was a problem hiding this comment.
I'm only wondering if this ends up being slower - you already have byte arrays, and now have to go through platform methods to read them?
Contributor
There was a problem hiding this comment.
+1.
It seems plausible that the new version will be faster, but it's probably a good idea to run a quick benchmark to confirm. There's a UTF8StringBenchmark linked from #19180 (comment) : maybe we could adapt that to work on byte arrays and do a quick before-and-after comparison to just to double check?
Edit: just to clarify: I noticed that this benchmark is also linked in the PR description. As Sean points out, I think the key difference in this PR is whether we're using getByte() versus directly accessing the on-heap byte array (in the linked UTF8String benchmark, both the old and new code were using getByte()).
Contributor
Author
There was a problem hiding this comment.
thank you @srowen and @JoshRosen for point out the difference. I follow the linked benchmark but add a new 512 byte slow benchmark which the first 511 bytes are same. The benchmark result shows it has no regression after this PR and has big benifits if the byte arrays have many same prefix.
Before this PR:
Java HotSpot(TM) 64-Bit Server VM 1.8.0_271-b09 on Mac OS X 10.16
Intel(R) Core(TM) i7-4770HQ CPU @ 2.20GHz
Byte Array compareTo: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
2-7 byte 800 861 70 81.9 12.2 1.0X
8-16 byte 810 878 59 80.9 12.4 1.0X
16-32 byte 804 887 40 81.5 12.3 1.0X
512-1024 byte 1050 1181 43 62.4 16.0 0.8X
512 byte slow 23593 23698 311 2.8 360.0 0.0X
2-7 byte 778 784 5 84.2 11.9 1.0X
After this PR:
Java HotSpot(TM) 64-Bit Server VM 1.8.0_271-b09 on Mac OS X 10.16
Intel(R) Core(TM) i7-4770HQ CPU @ 2.20GHz
Byte Array compareTo: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
2-7 byte 425 471 24 154.2 6.5 1.0X
8-16 byte 751 814 40 87.2 11.5 0.5X
16-32 byte 789 842 42 83.1 12.0 0.5X
512-1024 byte 1038 1175 193 63.1 15.8 0.4X
512 byte slow 3419 3924 NaN 19.2 52.2 0.1X
2-7 byte 421 424 2 155.6 6.4 1.0X
|
Kubernetes integration test starting |
|
Kubernetes integration test status failure |
|
Test build #144394 has finished for PR 34310 at commit
|
srowen
approved these changes
Oct 19, 2021
JoshRosen
reviewed
Oct 20, 2021
| benchmark.addCase("16-32 byte")(compareBinary(dataMedium)) | ||
| benchmark.addCase("512-1024 byte")(compareBinary(dataLarge)) | ||
| benchmark.addCase("512 byte slow")(compareBinary(dataLargeSlow)) | ||
| benchmark.addCase("2-7 byte")(compareBinary(dataTiny)) |
Contributor
There was a problem hiding this comment.
It looks like this this case is listed twice. Maybe drop this line?
Contributor
Author
There was a problem hiding this comment.
The first benchmark case may run slower than the latter due to the JIT optimization and this case has small size which can be done in a short time that would be more likely affected.
So I also keep it running twice in case this issue.
JoshRosen
reviewed
Oct 20, 2021
| byte array comparisons | ||
| ================================================================================================ | ||
|
|
||
| Java HotSpot(TM) 64-Bit Server VM 11.0.12+8-LTS-237 on Mac OS X 11.5 |
Contributor
There was a problem hiding this comment.
Can you run the benchmarks in GitHub Actions according to the instructions at https://spark.apache.org/developer-tools.html#github-workflow-benchmarks and then include those results in this PR in place of these ones? This helps to ensure that checked-in benchmark results come from a consistent environment.
Contributor
Author
There was a problem hiding this comment.
thank you @JoshRosen , it's really a good tool for me. Updated the benchmark result from GA.
Contributor
Author
There was a problem hiding this comment.
I also add the benchmark tool guide in pull request template, #34349
|
Kubernetes integration test starting |
|
Kubernetes integration test status failure |
|
Test build #144483 has finished for PR 34310 at commit
|
srowen
reviewed
Oct 21, 2021
| @@ -0,0 +1,16 @@ | |||
| ================================================================================================ | |||
Member
There was a problem hiding this comment.
Do we have 'before' numbers for these? you don't need to include them just want to verify that it also seemed to show an improvement like your local laptop one did
Contributor
Author
There was a problem hiding this comment.
Here is the old code path benchmark result:
JDK8
================================================================================================
byte array comparisons
================================================================================================
OpenJDK 64-Bit Server VM 1.8.0_312-b07 on Linux 5.8.0-1042-azure
Intel(R) Xeon(R) CPU E5-2673 v3 @ 2.40GHz
Byte Array compareTo: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
2-7 byte 799 836 24 82.0 12.2 1.0X
8-16 byte 832 906 32 78.8 12.7 1.0X
16-32 byte 812 854 28 80.7 12.4 1.0X
512-1024 byte 1057 1088 20 62.0 16.1 0.8X
512 byte slow 24628 26054 NaN 2.7 375.8 0.0X
2-7 byte 811 849 23 80.8 12.4 1.0X
JDK11
================================================================================================
byte array comparisons
================================================================================================
OpenJDK 64-Bit Server VM 11.0.13+8-LTS on Linux 5.8.0-1042-azure
Intel(R) Xeon(R) CPU E5-2673 v4 @ 2.30GHz
Byte Array compareTo: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
2-7 byte 771 812 28 85.0 11.8 1.0X
8-16 byte 839 857 13 78.1 12.8 0.9X
16-32 byte 898 926 17 73.0 13.7 0.9X
512-1024 byte 1141 1189 23 57.4 17.4 0.7X
512 byte slow 40124 40689 495 1.6 612.2 0.0X
2-7 byte 827 847 14 79.3 12.6 0.9X
Contributor
Author
There was a problem hiding this comment.
It shows we still have the benefits with GA env.
Contributor
Author
There was a problem hiding this comment.
ah, I just notice the env of GA is still different. The two benchmark result based on:
Intel(R) Xeon(R) Platinum 8171M CPU @ 2.60GHz
Intel(R) Xeon(R) CPU E5-2673 v4 @ 2.30GHz
Member
There was a problem hiding this comment.
I'm inclined to believe it is a win based on your first benchmark. Is there any easy way to run before/after on these Xeons, or is that hard?
Contributor
Author
There was a problem hiding this comment.
I compared the two code path within one patch, and here is the result.
JDK8:
================================================================================================
byte array comparisons
================================================================================================
OpenJDK 64-Bit Server VM 1.8.0_312-b07 on Linux 5.8.0-1042-azure
Intel(R) Xeon(R) CPU E5-2673 v4 @ 2.30GHz
Byte Array compare offHeap: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
2-7 byte 636 661 14 103.0 9.7 1.0X
8-16 byte 1067 1112 21 61.4 16.3 0.6X
16-32 byte 1226 1352 98 53.4 18.7 0.5X
512-1024 byte 1803 1916 46 36.3 27.5 0.4X
512 byte slow 4343 4662 171 15.1 66.3 0.1X
2-7 byte 1075 1119 26 61.0 16.4 0.6X
OpenJDK 64-Bit Server VM 1.8.0_312-b07 on Linux 5.8.0-1042-azure
Intel(R) Xeon(R) CPU E5-2673 v4 @ 2.30GHz
Byte Array compare onHeap: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
2-7 byte 1511 1570 30 43.4 23.1 1.0X
8-16 byte 1522 1564 27 43.1 23.2 1.0X
16-32 byte 1426 1554 36 46.0 21.8 1.1X
512-1024 byte 2080 2198 86 31.5 31.7 0.7X
512 byte slow 28498 29222 410 2.3 434.9 0.1X
2-7 byte 1382 1485 61 47.4 21.1 1.1X
================================================================================================
byte array comparisons
================================================================================================
OpenJDK 64-Bit Server VM 11.0.13+8-LTS on Linux 5.8.0-1042-azure
Intel(R) Xeon(R) CPU E5-2673 v3 @ 2.40GHz
Byte Array compare offHeap: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
2-7 byte 720 777 21 91.0 11.0 1.0X
8-16 byte 1077 1138 32 60.8 16.4 0.7X
16-32 byte 1347 1463 84 48.7 20.5 0.5X
512-1024 byte 1898 1989 40 34.5 29.0 0.4X
512 byte slow 4621 4878 168 14.2 70.5 0.2X
2-7 byte 1062 1133 28 61.7 16.2 0.7X
OpenJDK 64-Bit Server VM 11.0.13+8-LTS on Linux 5.8.0-1042-azure
Intel(R) Xeon(R) CPU E5-2673 v3 @ 2.40GHz
Byte Array compare onHeap: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
2-7 byte 1377 1471 37 47.6 21.0 1.0X
8-16 byte 1398 1475 38 46.9 21.3 1.0X
16-32 byte 1452 1547 47 45.2 22.1 0.9X
512-1024 byte 1826 1953 55 35.9 27.9 0.8X
512 byte slow 45883 47146 NaN 1.4 700.1 0.0X
2-7 byte 1401 1484 39 46.8 21.4 1.0X
srowen
approved these changes
Oct 24, 2021
Member
|
Merged to master |
Contributor
Author
|
thank you @srowen and @JoshRosen ! |
Member
|
late LGTM |
wangyum
pushed a commit
that referenced
this pull request
May 26, 2023
* [SPARK-36992][SQL] Improve byte array sort perf by unify getPrefix function of UTF8String and ByteArray ### What changes were proposed in this pull request? Unify the getPrefix function of `UTF8String` and `ByteArray`. ### Why are the changes needed? When execute sort operator, we first compare the prefix. However the getPrefix function of byte array is slow. We use first 8 bytes as the prefix, so at most we will call 8 times with `Platform.getByte` which is slower than call once with `Platform.getInt` or `Platform.getLong`. ### Does this PR introduce _any_ user-facing change? no ### How was this patch tested? pass `org.apache.spark.util.collection.unsafe.sort.PrefixComparatorsSuite` Closes #34267 from ulysses-you/binary-prefix. Authored-by: ulysses-you <[email protected]> Signed-off-by: Sean Owen <[email protected]> * [SPARK-37037][SQL] Improve byte array sort by unify compareTo function of UTF8String and ByteArray ### What changes were proposed in this pull request? Unify the compare function of `UTF8String` and `ByteArray`. ### Why are the changes needed? `BinaryType` use `TypeUtils.compareBinary` to compare two byte array, however it's slow since it compares byte array using unsigned int comparison byte by bye. We can compare them using `Platform.getLong` with unsigned long comparison if they have more than 8 bytes. And here is some histroy about this `TODO` https://github.com/apache/spark/pull/6755/files#r32197461 The benchmark result should be same with `UTF8String`, can be found in #19180 (#19180 (comment)) ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? Move test from `TypeUtilsSuite` to `ByteArraySuite` Closes #34310 from ulysses-you/SPARK-37037. Authored-by: ulysses-you <[email protected]> Signed-off-by: Sean Owen <[email protected]> * [SPARK-37341][SQL] Avoid unnecessary buffer and copy in full outer sort merge join ### What changes were proposed in this pull request? FULL OUTER sort merge join (non-code-gen path) [copies join keys and buffers input rows, even when rows from both sides do not have matched keys](https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/execution/joins/SortMergeJoinExec.scala#L1637-L1641). This is unnecessary, as we can just output the row with smaller join keys, and only buffer when both sides have matched keys. This would save us from unnecessary copy and buffer, when both join sides have a lot of rows not matched with each other. ### Why are the changes needed? Improve query performance for FULL OUTER sort merge join when code-gen is disabled. This would benefit query when both sides have a lot of rows not matched, and join key is big in terms of size (e.g. string type). Example micro benchmark: ``` def sortMergeJoin(): Unit = { val N = 2 << 20 codegenBenchmark("sort merge join", N) { val df1 = spark.range(N).selectExpr(s"cast(id * 15485863 as string) as k1") val df2 = spark.range(N).selectExpr(s"cast(id * 15485867 as string) as k2") val df = df1.join(df2, col("k1") === col("k2"), "full_outer") assert(df.queryExecution.sparkPlan.find(_.isInstanceOf[SortMergeJoinExec]).isDefined) df.noop() } } ``` Seeing run-time improvement over 60%: ``` Running benchmark: sort merge join Running case: sort merge join without optimization Stopped after 5 iterations, 10026 ms Running case: sort merge join with optimization Stopped after 5 iterations, 5954 ms Java HotSpot(TM) 64-Bit Server VM 1.8.0_181-b13 on Mac OS X 10.16 Intel(R) Core(TM) i9-9980HK CPU 2.40GHz sort merge join: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative ------------------------------------------------------------------------------------------------------------------------ sort merge join without optimization 1807 2005 157 1.2 861.4 1.0X sort merge join with optimization 1135 1191 62 1.8 541.1 1.6X ``` ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? Existing unit tests e.g. `OuterJoinSuite.scala`. Closes #34612 from c21/smj-fix. Authored-by: Cheng Su <[email protected]> Signed-off-by: Wenchen Fan <[email protected]> * [SPARK-37447][SQL] Cache LogicalPlan.isStreaming() result in a lazy val ### What changes were proposed in this pull request? This PR adds caching to `LogicalPlan.isStreaming()`: the default implementation's result will now be cached in a `private lazy val`. ### Why are the changes needed? This improves the performance of the `DeduplicateRelations` analyzer rule. The default implementation of `isStreaming` recursively visits every node in the tree. `DeduplicateRelations.renewDuplicatedRelations` is recursively invoked on every node in the tree and each invocation calls `isStreaming`. This leads to `O(n^2)` invocations of `isStreaming` on leaf nodes. Caching `isStreaming` avoids this performance problem. ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? Correctness should be covered by existing tests. This significantly improved `DeduplicateRelations` performance in local microbenchmarking with large query plans (~20% reduction in that rule's runtime in one of my tests). Closes #34691 from JoshRosen/cache-LogicalPlan.isStreaming. Authored-by: Josh Rosen <[email protected]> Signed-off-by: Wenchen Fan <[email protected]> * [SPARK-37530][CORE] Spark reads many paths very slow though newAPIHadoopFile ### What changes were proposed in this pull request? Same as #18441, we parallelize FileInputFormat.listStatus for newAPIHadoopFile ### Why are the changes needed?  Spark can be slow when accessing external storage at driver side, improve perf by parallelizing ### Does this PR introduce _any_ user-facing change? no ### How was this patch tested? passing GA Closes #34792 from yaooqinn/SPARK-37530. Authored-by: Kent Yao <[email protected]> Signed-off-by: Kent Yao <[email protected]> * [SPARK-37592][SQL] Improve performance of `JoinSelection` When I reading the implement of AQE, I find the process select join with hint exists a lot cumbersome code. The join hint has a relatively high learning curve for users, so the SQL not contains join hint in more cases. Improve performance of `JoinSelection` 'No'. Just change the inner implement. Jenkins test. Closes #34844 from beliefer/SPARK-37592-new. Authored-by: Jiaan Geng <[email protected]> Signed-off-by: Wenchen Fan <[email protected]> * [SPARK-37646][SQL] Avoid touching Scala reflection APIs in the lit function ### What changes were proposed in this pull request? This PR proposes to avoid touching Scala reflection APIs in the lit function. ### Why are the changes needed? Currently `lit` calls `typedlit[Any]` and touches Scala reflection APIs unnecessarily. As Scala reflection APIs touch multiple global locks and they are pretty slow when the parallelism is pretty high. This PR inlines `typedlit` to `lit` and replaces `Literal.create` with `Literal.apply` to avoid touching Scala reflection APIs. There is no behavior change. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? - New unit tests. - Manually ran the test in https://issues.apache.org/jira/browse/SPARK-37646 and saw no difference between `new Column(Literal(0L))` and `lit(0L)`. Closes #34901 from zsxwing/SPARK-37646. Lead-authored-by: Shixiong Zhu <[email protected]> Co-authored-by: Shixiong Zhu <[email protected]> Signed-off-by: Hyukjin Kwon <[email protected]> * [SPARK-37689][SQL] Expand should be supported in PropagateEmptyRelation We meet a case that when there is a empty relation, HashAggregateExec still triggered to execute and return an empty result. It's not necessary.  It's caused by there is an `Expand(EmptyLocalRelation())`, and it's not propagated, this pr support propagate `Expand` with empty LocalRelation Avoid unnecessary execution. No Added UT Closes #34954 from AngersZhuuuu/SPARK-37689. Authored-by: Angerszhuuuu <[email protected]> Signed-off-by: Wenchen Fan <[email protected]> * [SPARK-36406][CORE] Avoid unnecessary file operations before delete a write failed file held by DiskBlockObjectWriter We always do file truncate operation before delete a write failed file held by `DiskBlockObjectWriter`, a typical process is as follows: ``` if (!success) { // This code path only happens if an exception was thrown above before we set success; // close our stuff and let the exception be thrown further writer.revertPartialWritesAndClose() if (file.exists()) { if (!file.delete()) { logWarning(s"Error deleting ${file}") } } } ``` The `revertPartialWritesAndClose` method will reverts writes that haven't been committed yet, but it doesn't seem necessary in the current scene. So this pr add a new method to `DiskBlockObjectWriter` named `closeAndDelete()`, the new method just revert write metrics and delete the write failed file. Avoid unnecessary file operations. Add a new method to `DiskBlockObjectWriter` named `closeAndDelete(). Pass the Jenkins or GitHub Action Closes #33628 from LuciferYang/SPARK-36406. Authored-by: yangjie01 <[email protected]> Signed-off-by: attilapiros <[email protected]> * [SPARK-37462][CORE] Avoid unnecessary calculating the number of outstanding fetch requests and RPCS Avoid unnecessary calculating the number of outstanding fetch requests and RPCS It is unnecessary to calculate the number of outstanding fetch requests and RPCS when the IdleStateEvent is not IDLE or the last request is not timeout. No. Exist unittests. Closes #34711 from weixiuli/SPARK-37462. Authored-by: weixiuli <[email protected]> Signed-off-by: Sean Owen <[email protected]> Co-authored-by: ulysses-you <[email protected]> Co-authored-by: Cheng Su <[email protected]> Co-authored-by: Josh Rosen <[email protected]> Co-authored-by: Kent Yao <[email protected]> Co-authored-by: Jiaan Geng <[email protected]> Co-authored-by: Shixiong Zhu <[email protected]> Co-authored-by: Shixiong Zhu <[email protected]> Co-authored-by: Angerszhuuuu <[email protected]> Co-authored-by: yangjie01 <[email protected]> Co-authored-by: weixiuli <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What changes were proposed in this pull request?
Unify the compare function of
UTF8StringandByteArray.Why are the changes needed?
BinaryTypeuseTypeUtils.compareBinaryto compare two byte array, however it's slow since it compares byte array using unsigned int comparison byte by bye.We can compare them using
Platform.getLongwith unsigned long comparison if they have more than 8 bytes. And here is some histroy about thisTODOhttps://github.com/apache/spark/pull/6755/files#r32197461The benchmark result should be same with
UTF8String, can be found in #19180 (#19180 (comment))Does this PR introduce any user-facing change?
No
How was this patch tested?
TypeUtilsSuitetoByteArraySuiteJDK8:
JDK11