c21 opened a new pull request #31736:
URL: https://github.com/apache/spark/pull/31736
<!--
Thanks for sending a pull request! Here are some tips for you:
1. If this is your first time, please read our contributor guidelines:
https://spark.apache.org/contributing.html
2. Ensure you have added or run the appropriate tests for your PR:
https://spark.apache.org/developer-tools.html
3. If the PR is unfinished, add '[WIP]' in your PR title, e.g.,
'[WIP][SPARK-XXXX] Your PR title ...'.
4. Be sure to keep the PR description updated to reflect all changes.
5. Please write your PR title to summarize what this PR proposes.
6. If possible, provide a concise example to reproduce the issue for a
faster review.
7. If you want to add a new configuration, please read the guideline first
for naming configurations in
'core/src/main/scala/org/apache/spark/internal/config/ConfigEntry.scala'.
-->
### What changes were proposed in this pull request?
<!--
Please clarify what changes you are proposing. The purpose of this section
is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR. See the examples below.
1. If you refactor some codes with changing classes, showing the class
hierarchy will help reviewers.
2. If you fix some SQL features, you can provide some references of other
DBMSes.
3. If there is design documentation, please add the link.
4. If there is a discussion in the mailing list, please add the link.
-->
`BroadcastNestedLoopJoinExec` does not have code-gen, and we can potentially
boost the CPU performance for this operator if we add code-gen for it.
https://databricks.com/blog/2017/02/16/processing-trillion-rows-per-second-single-machine-can-nested-loop-joins-fast.html
also showed the evidence in one fork.
The codegen for `BroadcastNestedLoopJoinExec` shared some code with
`HashJoin`, and the interface `JoinCodegenSupport` is created to hold those
common logic. This PR is only supporting inner and cross join. Other join types
will be added later in followup PRs.
Example query and generated code:
```
val df1 = spark.range(4).select($"id".as("k1"))
val df2 = spark.range(3).select($"id".as("k2"))
df1.join(df2, $"k1" + 1 =!= $"k2").explain("codegen")
```
```
== Subtree 2 / 2 (maxMethodCodeSize:282; maxConstantPoolSize:203(0.31%
used); numInnerClasses:0) ==
*(2) BroadcastNestedLoopJoin BuildRight, Inner, NOT ((k1#2L + 1) = k2#6L)
:- *(2) Project [id#0L AS k1#2L]
: +- *(2) Range (0, 4, step=1, splits=2)
+- BroadcastExchange IdentityBroadcastMode, [id=#22]
+- *(1) Project [id#4L AS k2#6L]
+- *(1) Range (0, 3, step=1, splits=2)
Generated code:
/* 001 */ public Object generate(Object[] references) {
/* 002 */ return new GeneratedIteratorForCodegenStage2(references);
/* 003 */ }
/* 004 */
/* 005 */ // codegenStageId=2
/* 006 */ final class GeneratedIteratorForCodegenStage2 extends
org.apache.spark.sql.execution.BufferedRowIterator {
/* 007 */ private Object[] references;
/* 008 */ private scala.collection.Iterator[] inputs;
/* 009 */ private boolean range_initRange_0;
/* 010 */ private long range_nextIndex_0;
/* 011 */ private TaskContext range_taskContext_0;
/* 012 */ private InputMetrics range_inputMetrics_0;
/* 013 */ private long range_batchEnd_0;
/* 014 */ private long range_numElementsTodo_0;
/* 015 */ private InternalRow[] bnlj_broadcastArray_0;
/* 016 */ private
org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter[]
range_mutableStateArray_0 = new
org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter[4];
/* 017 */
/* 018 */ public GeneratedIteratorForCodegenStage2(Object[] references) {
/* 019 */ this.references = references;
/* 020 */ }
/* 021 */
/* 022 */ public void init(int index, scala.collection.Iterator[] inputs) {
/* 023 */ partitionIndex = index;
/* 024 */ this.inputs = inputs;
/* 025 */
/* 026 */ range_taskContext_0 = TaskContext.get();
/* 027 */ range_inputMetrics_0 =
range_taskContext_0.taskMetrics().inputMetrics();
/* 028 */ range_mutableStateArray_0[0] = new
org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(1, 0);
/* 029 */ range_mutableStateArray_0[1] = new
org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(1, 0);
/* 030 */ range_mutableStateArray_0[2] = new
org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(1, 0);
/* 031 */ bnlj_broadcastArray_0 = (InternalRow[])
((org.apache.spark.broadcast.TorrentBroadcast) references[1] /* broadcastTerm
*/).value();
/* 032 */ range_mutableStateArray_0[3] = new
org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(2, 0);
/* 033 */
/* 034 */ }
/* 035 */
/* 036 */ private void bnlj_doConsume_0(long bnlj_expr_0_0) throws
java.io.IOException {
/* 037 */ int bnlj_arrayIndex_0 = 0;
/* 038 */ UnsafeRow bnlj_buildRow_0;
/* 039 */ while (bnlj_arrayIndex_0 < bnlj_broadcastArray_0.length) {
/* 040 */ bnlj_buildRow_0 = (UnsafeRow)
bnlj_broadcastArray_0[bnlj_arrayIndex_0];
/* 041 */
/* 042 */ long bnlj_value_1 = bnlj_buildRow_0.getLong(0);
/* 043 */
/* 044 */ long bnlj_value_4 = -1L;
/* 045 */
/* 046 */ bnlj_value_4 = bnlj_expr_0_0 + 1L;
/* 047 */
/* 048 */ boolean bnlj_value_3 = false;
/* 049 */ bnlj_value_3 = bnlj_value_4 == bnlj_value_1;
/* 050 */ boolean bnlj_value_2 = false;
/* 051 */ bnlj_value_2 = !(bnlj_value_3);
/* 052 */ if (!(false || !bnlj_value_2))
/* 053 */ {
/* 054 */ ((org.apache.spark.sql.execution.metric.SQLMetric)
references[2] /* numOutputRows */).add(1);
/* 055 */
/* 056 */ range_mutableStateArray_0[3].reset();
/* 057 */
/* 058 */ range_mutableStateArray_0[3].write(0, bnlj_expr_0_0);
/* 059 */
/* 060 */ range_mutableStateArray_0[3].write(1, bnlj_value_1);
/* 061 */ append((range_mutableStateArray_0[3].getRow()).copy());
/* 062 */
/* 063 */ }
/* 064 */ bnlj_arrayIndex_0 += 1;
/* 065 */ }
/* 066 */
/* 067 */ }
/* 068 */
/* 069 */ private void initRange(int idx) {
/* 070 */ java.math.BigInteger index = java.math.BigInteger.valueOf(idx);
/* 071 */ java.math.BigInteger numSlice =
java.math.BigInteger.valueOf(2L);
/* 072 */ java.math.BigInteger numElement =
java.math.BigInteger.valueOf(4L);
/* 073 */ java.math.BigInteger step = java.math.BigInteger.valueOf(1L);
/* 074 */ java.math.BigInteger start = java.math.BigInteger.valueOf(0L);
/* 075 */ long partitionEnd;
/* 076 */
/* 077 */ java.math.BigInteger st =
index.multiply(numElement).divide(numSlice).multiply(step).add(start);
/* 078 */ if (st.compareTo(java.math.BigInteger.valueOf(Long.MAX_VALUE))
> 0) {
/* 079 */ range_nextIndex_0 = Long.MAX_VALUE;
/* 080 */ } else if
(st.compareTo(java.math.BigInteger.valueOf(Long.MIN_VALUE)) < 0) {
/* 081 */ range_nextIndex_0 = Long.MIN_VALUE;
/* 082 */ } else {
/* 083 */ range_nextIndex_0 = st.longValue();
/* 084 */ }
/* 085 */ range_batchEnd_0 = range_nextIndex_0;
/* 086 */
/* 087 */ java.math.BigInteger end =
index.add(java.math.BigInteger.ONE).multiply(numElement).divide(numSlice)
/* 088 */ .multiply(step).add(start);
/* 089 */ if
(end.compareTo(java.math.BigInteger.valueOf(Long.MAX_VALUE)) > 0) {
/* 090 */ partitionEnd = Long.MAX_VALUE;
/* 091 */ } else if
(end.compareTo(java.math.BigInteger.valueOf(Long.MIN_VALUE)) < 0) {
/* 092 */ partitionEnd = Long.MIN_VALUE;
/* 093 */ } else {
/* 094 */ partitionEnd = end.longValue();
/* 095 */ }
/* 096 */
/* 097 */ java.math.BigInteger startToEnd =
java.math.BigInteger.valueOf(partitionEnd).subtract(
/* 098 */ java.math.BigInteger.valueOf(range_nextIndex_0));
/* 099 */ range_numElementsTodo_0 = startToEnd.divide(step).longValue();
/* 100 */ if (range_numElementsTodo_0 < 0) {
/* 101 */ range_numElementsTodo_0 = 0;
/* 102 */ } else if
(startToEnd.remainder(step).compareTo(java.math.BigInteger.valueOf(0L)) != 0) {
/* 103 */ range_numElementsTodo_0++;
/* 104 */ }
/* 105 */ }
/* 106 */
/* 107 */ protected void processNext() throws java.io.IOException {
/* 108 */ // initialize Range
/* 109 */ if (!range_initRange_0) {
/* 110 */ range_initRange_0 = true;
/* 111 */ initRange(partitionIndex);
/* 112 */ }
/* 113 */
/* 114 */ while (true) {
/* 115 */ if (range_nextIndex_0 == range_batchEnd_0) {
/* 116 */ long range_nextBatchTodo_0;
/* 117 */ if (range_numElementsTodo_0 > 1000L) {
/* 118 */ range_nextBatchTodo_0 = 1000L;
/* 119 */ range_numElementsTodo_0 -= 1000L;
/* 120 */ } else {
/* 121 */ range_nextBatchTodo_0 = range_numElementsTodo_0;
/* 122 */ range_numElementsTodo_0 = 0;
/* 123 */ if (range_nextBatchTodo_0 == 0) break;
/* 124 */ }
/* 125 */ range_batchEnd_0 += range_nextBatchTodo_0 * 1L;
/* 126 */ }
/* 127 */
/* 128 */ int range_localEnd_0 = (int)((range_batchEnd_0 -
range_nextIndex_0) / 1L);
/* 129 */ for (int range_localIdx_0 = 0; range_localIdx_0 <
range_localEnd_0; range_localIdx_0++) {
/* 130 */ long range_value_0 = ((long)range_localIdx_0 * 1L) +
range_nextIndex_0;
/* 131 */
/* 132 */ // common sub-expressions
/* 133 */
/* 134 */ bnlj_doConsume_0(range_value_0);
/* 135 */
/* 136 */ if (shouldStop()) {
/* 137 */ range_nextIndex_0 = range_value_0 + 1L;
/* 138 */ ((org.apache.spark.sql.execution.metric.SQLMetric)
references[0] /* numOutputRows */).add(range_localIdx_0 + 1);
/* 139 */ range_inputMetrics_0.incRecordsRead(range_localIdx_0 +
1);
/* 140 */ return;
/* 141 */ }
/* 142 */
/* 143 */ }
/* 144 */ range_nextIndex_0 = range_batchEnd_0;
/* 145 */ ((org.apache.spark.sql.execution.metric.SQLMetric)
references[0] /* numOutputRows */).add(range_localEnd_0);
/* 146 */ range_inputMetrics_0.incRecordsRead(range_localEnd_0);
/* 147 */ range_taskContext_0.killTaskIfInterrupted();
/* 148 */ }
/* 149 */ }
/* 150 */
/* 151 */ }
```
### Why are the changes needed?
<!--
Please clarify why the changes are needed. For instance,
1. If you propose a new API, clarify the use case for a new API.
2. If you fix a bug, you can clarify why it is a bug.
-->
Improve query CPU performance. Added a micro benchmark query in
`JoinBenchmark.scala`.
Saw 70% of run time improvement:
```
Running benchmark: broadcast nested loop join
Running case: broadcast nested loop join wholestage off
Stopped after 2 iterations, 48367 ms
Running case: broadcast nested loop join wholestage on
Stopped after 5 iterations, 74502 ms
Java HotSpot(TM) 64-Bit Server VM 1.8.0_181-b13 on Mac OS X 10.15.7
Intel(R) Core(TM) i9-9980HK CPU @ 2.40GHz
broadcast nested loop join: Best Time(ms) Avg Time(ms)
Stdev(ms) Rate(M/s) Per Row(ns) Relative
-------------------------------------------------------------------------------------------------------------------------
broadcast nested loop join wholestage off 24130 24184
76 0.9 1150.6 1.0X
broadcast nested loop join wholestage on 14589 14900
220 1.4 695.7 1.7X
```
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such as
the documentation fix.
If yes, please clarify the previous behavior and the change this PR proposes
- provide the console output, description and/or an example to show the
behavior difference if possible.
If possible, please also clarify if this is a user-facing change compared to
the released Spark versions or within the unreleased branches such as master.
If no, write 'No'.
-->
No.
### How was this patch tested?
<!--
If tests were added, say they were added here. Please make sure to add some
test cases that check the changes thoroughly including negative and positive
cases if possible.
If it was tested in a way different from regular unit tests, please clarify
how you tested step by step, ideally copy and paste-able, so that other
reviewers can test and check, and descendants can verify in the future.
If tests were not added, please describe why they were not added and/or why
it was difficult to add.
-->
Added unit test in `WholeStageCodegenSuite.scala`, and existing unit tests
for `BroadcastNestedLoopJoinExec`.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
