Re: RFR: 8334342: Add MergeStore JMH benchmarks [v5]
On Tue, 18 Jun 2024 01:09:49 GMT, Shaojin Wen wrote: >> Shaojin Wen has updated the pull request with a new target base due to a >> merge or a rebase. The incremental webrev excludes the unrelated changes >> brought in by the merge/rebase. The pull request contains 13 additional >> commits since the last revision: >> >> - Merge remote-tracking branch 'upstream/master' into merge_store_bench >> - bug fix for `putChars4C` >> - bug fix for `putChars4C` and `putChars4S` >> - use VarHandler CHAR_L & CHAR_B >> - copyright >> - bug fix for putIntBU >> - add cases for `getChar` & `putChar` >> - code format >> - add `setIntRL` & `setIntRLU` >> - add comments >> - ... and 3 more: https://git.openjdk.org/jdk/compare/322c6e37...4c9b9418 > > I re-ran the performance test based on WebRevs 04: > [Full](https://webrevs.openjdk.org/?repo=jdk&pr=19734&range=04) - > [Incremental](https://webrevs.openjdk.org/?repo=jdk&pr=19734&range=03-04) > ([4c9b9418](https://git.openjdk.org/jdk/pull/19734/files/4c9b9418fc4a95504867b6019b3e94605917f747)) > . > > # 1. Cases MergeStore does not work > @eme64 > I found `putChars4BV` and `putChars4LV` two cases MergeStore didn't work, if > support can be enhanced, it would be useful for people using VarHandle. > > > putChars4BV > putChars4LV > > > I also found that the performance of the case using VarHandle is particularly > good. Why? For example: > > setIntBV > setIntLV > setLongBV > setLongLV > > > # 2. Performance numbers > The names of these cases have the following B/L/V/U suffixes, which are: > > B BigEndian > L LittleEndian > V VarHandle > U Unsafe > R reverseBytes > C Unsafe.getChar & putChar > S Unsafe.getShort & putShort > > > ## 2.1 MacBook M1 Pro (aarch64) > > BenchmarkMode Cnt ScoreError Units > MergeStoreBench.getCharB avgt 15 5340.200 ? 7.038 ns/op > MergeStoreBench.getCharBUavgt 15 5482.163 ? 7.922 ns/op > MergeStoreBench.getCharBVavgt 15 5074.165 ? 6.759 ns/op > MergeStoreBench.getCharC avgt 15 5051.763 ? 6.552 ns/op > MergeStoreBench.getCharL avgt 15 5374.464 ? 9.783 ns/op > MergeStoreBench.getCharLUavgt 15 5487.532 ? 6.368 ns/op > MergeStoreBench.getCharLVavgt 15 5071.263 ? 9.717 ns/op > MergeStoreBench.getIntB avgt 15 6277.984 ? 6.284 ns/op > MergeStoreBench.getIntBU avgt 15 5232.984 ? 10.384 ns/op > MergeStoreBench.getIntBV avgt 15 1206.264 ? 1.193 ns/op > MergeStoreBench.getIntL avgt 15 6172.779 ? 1.962 ns/op > MergeStoreBench.getIntLU avgt 15 5157.317 ? 16.077 ns/op > MergeStoreBench.getIntLV avgt 15 2558.110 ? 3.402 ns/op > MergeStoreBench.getIntRB avgt 15 6889.916 ? 36.955 ns/op > MergeStoreBench.getIntRBUavgt 15 5769.950 ? 11.499 ns/op > MergeStoreBench.getIntRL avgt 15 6625.605 ? 10.662 ns/op > MergeStoreBench.getIntRLUavgt 15 5746.742 ? 11.945 ns/op > MergeStoreBench.getIntRU avgt 15 2544.586 ? 2.769 ns/op > MergeStoreBench.getIntU avgt 15 2541.119 ? 3.252 ns/op > MergeStoreBench.getLongB avgt 15 12098.129 ? 31.451 ns/op > MergeStoreBench.getLongBUavgt 15 9760.621 ? 16.427 ns/op > MergeStoreBench.getLongBVavgt 15 2593.635 ? 4.698 ns/op > MergeStoreBench.getLongL avgt 15 12031.065 ? 19.820 ns/op > Merge... @wenshao what is the state of this PR? - PR Comment: https://git.openjdk.org/jdk/pull/19734#issuecomment-2247964317
Re: RFR: 8334342: Add MergeStore JMH benchmarks [v5]
On Mon, 17 Jun 2024 23:02:47 GMT, Shaojin Wen wrote: >> [8318446](https://github.com/openjdk/jdk/pull/16245) brings MergeStore. We >> need a JMH Benchmark to evaluate the performance of various batch operations >> and the effect of MergeStore. > > Shaojin Wen has updated the pull request with a new target base due to a > merge or a rebase. The incremental webrev excludes the unrelated changes > brought in by the merge/rebase. The pull request contains 13 additional > commits since the last revision: > > - Merge remote-tracking branch 'upstream/master' into merge_store_bench > - bug fix for `putChars4C` > - bug fix for `putChars4C` and `putChars4S` > - use VarHandler CHAR_L & CHAR_B > - copyright > - bug fix for putIntBU > - add cases for `getChar` & `putChar` > - code format > - add `setIntRL` & `setIntRLU` > - add comments > - ... and 3 more: https://git.openjdk.org/jdk/compare/d4942ac0...4c9b9418 I re-ran the performance test based on WebRevs 04: [Full](https://webrevs.openjdk.org/?repo=jdk&pr=19734&range=04) - [Incremental](https://webrevs.openjdk.org/?repo=jdk&pr=19734&range=03-04) ([4c9b9418](https://git.openjdk.org/jdk/pull/19734/files/4c9b9418fc4a95504867b6019b3e94605917f747)) . # 1. Cases MergeStore does not work @eme64 I found `putChars4BV` and `putChars4LV` to be two cases where MergeStore didn't work, if support can be enhanced, it would be useful for people using VarHandle. putChars4BV putChars4LV I also found that the performance of the case using VarHandle is particularly good. Why? For example: setIntBV setIntLV setLongBV setLongLV # 2. Performance numbers The names of these cases have the following B/L/V/U suffixes, which are: B BigEndian L LittleEndian V VarHandle U Unsafe R reverseBytes C Unsafe.getChar & putChar S Unsafe.getShort & putShort ## 2.1 MacBook M1 Pro (aarch64) BenchmarkMode Cnt ScoreError Units MergeStoreBench.getCharB avgt 15 5340.200 ? 7.038 ns/op MergeStoreBench.getCharBUavgt 15 5482.163 ? 7.922 ns/op MergeStoreBench.getCharBVavgt 15 5074.165 ? 6.759 ns/op MergeStoreBench.getCharC avgt 15 5051.763 ? 6.552 ns/op MergeStoreBench.getCharL avgt 15 5374.464 ? 9.783 ns/op MergeStoreBench.getCharLUavgt 15 5487.532 ? 6.368 ns/op MergeStoreBench.getCharLVavgt 15 5071.263 ? 9.717 ns/op MergeStoreBench.getIntB avgt 15 6277.984 ? 6.284 ns/op MergeStoreBench.getIntBU avgt 15 5232.984 ? 10.384 ns/op MergeStoreBench.getIntBV avgt 15 1206.264 ? 1.193 ns/op MergeStoreBench.getIntL avgt 15 6172.779 ? 1.962 ns/op MergeStoreBench.getIntLU avgt 15 5157.317 ? 16.077 ns/op MergeStoreBench.getIntLV avgt 15 2558.110 ? 3.402 ns/op MergeStoreBench.getIntRB avgt 15 6889.916 ? 36.955 ns/op MergeStoreBench.getIntRBUavgt 15 5769.950 ? 11.499 ns/op MergeStoreBench.getIntRL avgt 15 6625.605 ? 10.662 ns/op MergeStoreBench.getIntRLUavgt 15 5746.742 ? 11.945 ns/op MergeStoreBench.getIntRU avgt 15 2544.586 ? 2.769 ns/op MergeStoreBench.getIntU avgt 15 2541.119 ? 3.252 ns/op MergeStoreBench.getLongB avgt 15 12098.129 ? 31.451 ns/op MergeStoreBench.getLongBUavgt 15 9760.621 ? 16.427 ns/op MergeStoreBench.getLongBVavgt 15 2593.635 ? 4.698 ns/op MergeStoreBench.getLongL avgt 15 12031.065 ? 19.820 ns/op MergeStoreBench.getLongLUavgt 15 9653.938 ? 18.372 ns/op MergeStoreBench.getLongLVavgt 15 2557.521 ? 3.338 ns/op MergeStoreBench.getLongRBavgt 15 12092.061 ? 18.026 ns/op MergeStoreBench.getLongRBU avgt 15 9763.489 ? 17.347 ns/op MergeStoreBench.getLongRLavgt 15 12027.686 ? 17.472 ns/op MergeStoreBench.getLongRLU avgt 15 9649.433 ? 8.384 ns/op MergeStoreBench.getLongRUavgt 15 2546.239 ? 2.088 ns/op MergeStoreBench.getLongU avgt 15 2539.762 ? 1.439 ns/op MergeStoreBench.putChars4B avgt 15 8487.381 ? 23.170 ns/op MergeStoreBench.putChars4BU avgt 15 3830.198 ? 7.083 ns/op MergeStoreBench.putChars4BV avgt 15 5154.819 ? 10.348 ns/op MergeStoreBench.putChars4C avgt 15 5162.766 ? 15.041 ns/op MergeStoreBench.putChars4L avgt 15 8381.231 ? 20.135 ns/op MergeStoreBench.putChars4LU avgt 15 3827.784 ? 3.163 ns/op MergeStoreBench.putChars4LV avgt 15 5151.508 ? 4.907 ns/op MergeStoreBench.putChars4S avgt 15 5152.123 ? 7.407 ns/op MergeStoreBench.setCharBSavgt 15 5317.319 ? 28.445 ns/op MergeStoreBench.setCharBVavgt 15 5175.400 ? 7.110 ns/op MergeStoreBench.setCharC avgt 15 5085.752 ? 6.222 ns/op MergeStoreBench.setCharLSavgt 15 5294.766 ? 9.742 ns/op MergeStoreBench.setCharLVavgt 15 5108.269 ? 6.692 ns/op MergeStoreBench.setIntB avgt 15 5095.236 ? 2.838 ns/op MergeStoreBench.setIntBU avgt 15 5097.007 ? 4.249 ns/op MergeStoreBench.setIntBV avgt
Re: RFR: 8334342: Add MergeStore JMH benchmarks [v5]
> [8318446](https://github.com/openjdk/jdk/pull/16245) brings MergeStore. We > need a JMH Benchmark to evaluate the performance of various batch operations > and the effect of MergeStore. Shaojin Wen has updated the pull request with a new target base due to a merge or a rebase. The incremental webrev excludes the unrelated changes brought in by the merge/rebase. The pull request contains 13 additional commits since the last revision: - Merge remote-tracking branch 'upstream/master' into merge_store_bench - bug fix for `putChars4C` - bug fix for `putChars4C` and `putChars4S` - use VarHandler CHAR_L & CHAR_B - copyright - bug fix for putIntBU - add cases for `getChar` & `putChar` - code format - add `setIntRL` & `setIntRLU` - add comments - ... and 3 more: https://git.openjdk.org/jdk/compare/fc9315ff...4c9b9418 - Changes: - all: https://git.openjdk.org/jdk/pull/19734/files - new: https://git.openjdk.org/jdk/pull/19734/files/1140bd81..4c9b9418 Webrevs: - full: https://webrevs.openjdk.org/?repo=jdk&pr=19734&range=04 - incr: https://webrevs.openjdk.org/?repo=jdk&pr=19734&range=03-04 Stats: 1440 lines in 77 files changed: 795 ins; 459 del; 186 mod Patch: https://git.openjdk.org/jdk/pull/19734.diff Fetch: git fetch https://git.openjdk.org/jdk.git pull/19734/head:pull/19734 PR: https://git.openjdk.org/jdk/pull/19734
