Sorry for the confusion. To explain, I use a local ann-benchmarks set-up that makes use of KnnGraphTester. It is a bit hacky and I accidentally included the warm-ups in the final timings. So the change to warm-up explains why we saw different results in our tests. This is great motivation to solidify and publish my local ann-benchmarks set-up so that it's not so fragile!
In summary, with your latest fix the recall and QPS look good to me -- I don't detect any regression between 9.3 and 9.4. Julie On Mon, Sep 19, 2022 at 3:45 PM Michael Sokolov <msoko...@gmail.com> wrote: > I'm confused, since warming should not be counted in the timings. Are you > saying that the recall was affected?? > > On Mon, Sep 19, 2022 at 6:12 PM Julie Tibshirani <juliet...@gmail.com> > wrote: > >> Using the ann-benchmarks framework, I still saw a similar regression as >> Mayya between 9.3 and 9.4. I investigated and found it was due to >> "KnnGraphTester to use KnnVectorQuery" ( >> https://github.com/apache/lucene/pull/796), specifically the change to >> the warm-up strategy. If I revert it, the results look exactly as expected. >> >> I guess we can keep an eye on the nightly benchmarks tomorrow to >> double-check there's no drop. It would also be nice to formalize the >> ann-benchmarks set-up and run it regularly (like we've discussed in >> https://github.com/apache/lucene/issues/10665). >> >> Julie >> >> On Mon, Sep 19, 2022 at 10:33 AM Michael Sokolov <msoko...@gmail.com> >> wrote: >> >>> Thanks for your speedy testing! I am observing comparable latencies >>> *when the index geometry (ie number of segments)* is unchanged. Agree we >>> can leave this for a later day. I'll proceed to cut 9.4 artifacts >>> >>> On Mon, Sep 19, 2022 at 11:02 AM Mayya Sharipova >>> <mayya.sharip...@elastic.co.invalid> wrote: >>> >>>> It would be great if you all are able to test again with >>>>> https://github.com/apache/lucene/pull/11781/ applied >>>> >>>> >>>> >>>> I ran the ann benchmarks with this change, and was happy to confirm >>>> that in my test recall with this PR is the same as in 9.3 branch, although >>>> QPS is lower, but we can investigate QPSs later. >>>> >>>> glove-100-angular M:16 efConstruction:100 >>>> 9.3 recall9.3 QPSthis PR recallthis PR QPS >>>> n_cands=10 0.620 2745.933 0.620 1675.500 >>>> n_cands=20 0.680 2288.665 0.680 1512.744 >>>> n_cands=40 0.746 1770.243 0.746 1040.240 >>>> n_cands=80 0.809 1226.738 0.809 695.236 >>>> n_cands=120 0.843 948.908 0.843 525.914 >>>> n_cands=200 0.878 671.781 0.878 351.529 >>>> n_cands=400 0.918 392.265 0.918 207.854 >>>> n_cands=600 0.937 282.403 0.937 144.311 >>>> n_cands=800 0.949 214.620 0.949 116.875 >>>> >>>> On Sun, Sep 18, 2022 at 6:25 PM Michael Sokolov <msoko...@gmail.com> >>>> wrote: >>>> >>>>> OK, I think I was wrong about latency having increased due to a change >>>>> in KnnGraphTester -- I did some testing there and couldn't reproduce. >>>>> There does seem to be a slight vector search latency increase, >>>>> possibly noise, but maybe due to the branching introduced to check >>>>> whether to do byte vs float operations? It would be a little >>>>> surprising if that were the case given the small number of branchings >>>>> compared to the number of multiplies in dot-product though. >>>>> >>>>> On Sun, Sep 18, 2022 at 3:25 PM Michael Sokolov <msoko...@gmail.com> >>>>> wrote: >>>>> > >>>>> > Thanks for the deep-dive Julie. I was able to reproduce the changing >>>>> > recall. I had introduced some bugs in the diversity checks (that may >>>>> > have partially canceled each other out? it's hard to understand what >>>>> > was happening in the buggy case) and posted a fix today >>>>> > https://github.com/apache/lucene/pull/11781. >>>>> > >>>>> > There are a couple of other outstanding issues I found while doing a >>>>> > bunch of git bisecting; >>>>> > >>>>> > I think we might have introduced a (test-only) performance regression >>>>> > in KnnGraphTester >>>>> > >>>>> > We may still be over-allocating the size of NeighborArray, leading to >>>>> > excessive segmentation? I wonder if we could avoid dynamic >>>>> > re-allocation there, and simply initialize every neighbor array to >>>>> > 2*M+1. >>>>> > >>>>> > While I don't think these are necessarily blockers, given that we are >>>>> > releasing HNSW improvements, it seems like we should address these, >>>>> > especially as the build-graph-on-index is one of the things we are >>>>> > releasing, and it is (may be?) impacted. I will see if I can put up a >>>>> > patch or two. >>>>> > >>>>> > It would be great if you all are able to test again with >>>>> > https://github.com/apache/lucene/pull/11781/ applied >>>>> > >>>>> > -Mike >>>>> > >>>>> > On Fri, Sep 16, 2022 at 11:07 AM Adrien Grand <jpou...@gmail.com> >>>>> wrote: >>>>> > > >>>>> > > Thank you Mike, I just backported the change. >>>>> > > >>>>> > > On Thu, Sep 15, 2022 at 6:32 PM Michael Sokolov < >>>>> msoko...@gmail.com> wrote: >>>>> > >> >>>>> > >> it looks like a small bug fix, we have had on main (and 9.x?) for >>>>> a >>>>> > >> while now and no test failures showed up, I guess. Should be OK to >>>>> > >> port. I plan to cut artifacts this weekend, or Monday at the >>>>> latest, >>>>> > >> but if you can do the backport today or tomorrow, that's fine by >>>>> me. >>>>> > >> >>>>> > >> On Thu, Sep 15, 2022 at 10:55 AM Adrien Grand <jpou...@gmail.com> >>>>> wrote: >>>>> > >> > >>>>> > >> > Mike, I'm tempted to backport >>>>> https://github.com/apache/lucene/pull/1068 to branch_9_4, which is a >>>>> bugfix that looks pretty safe to me. What do you think? >>>>> > >> > >>>>> > >> > On Tue, Sep 13, 2022 at 4:11 PM Mayya Sharipova < >>>>> mayya.sharip...@elastic.co.invalid> wrote: >>>>> > >> >> >>>>> > >> >> Thanks for running more tests, Michael. >>>>> > >> >> It is encouraging that you saw a similar performance between >>>>> 9.3 and 9.4. I will also run more tests with different parameters. >>>>> > >> >> >>>>> > >> >> On Tue, Sep 13, 2022 at 9:30 AM Michael Sokolov < >>>>> msoko...@gmail.com> wrote: >>>>> > >> >>> >>>>> > >> >>> As a follow-up, I ran a test using the same parameters as >>>>> above, only >>>>> > >> >>> changing M=200 to M=16. This did result in a single segment >>>>> in both >>>>> > >> >>> cases (9.3, 9.4) and the performance was pretty similar; >>>>> within noise >>>>> > >> >>> I think. The main difference I saw was that the 9.3 index was >>>>> written >>>>> > >> >>> using CFS: >>>>> > >> >>> >>>>> > >> >>> 9.4: >>>>> > >> >>> recall latency nDoc fanout maxConn beamWidth >>>>> visited index ms >>>>> > >> >>> 0.755 1.36 1000000 100 16 100 200 >>>>> 891402 1.00 >>>>> > >> >>> post-filter >>>>> > >> >>> -rw-r--r-- 1 sokolovm amazon 382M Sep 13 13:06 >>>>> > >> >>> _0_Lucene94HnswVectorsFormat_0.vec >>>>> > >> >>> -rw-r--r-- 1 sokolovm amazon 262K Sep 13 13:06 >>>>> > >> >>> _0_Lucene94HnswVectorsFormat_0.vem >>>>> > >> >>> -rw-r--r-- 1 sokolovm amazon 131M Sep 13 13:06 >>>>> > >> >>> _0_Lucene94HnswVectorsFormat_0.vex >>>>> > >> >>> >>>>> > >> >>> 9.3: >>>>> > >> >>> recall latency nDoc fanout maxConn beamWidth >>>>> visited index ms >>>>> > >> >>> 0.775 1.34 1000000 100 16 100 4033 977043 >>>>> > >> >>> rw-r--r-- 1 sokolovm amazon 297 Sep 13 13:26 _0.cfe >>>>> > >> >>> -rw-r--r-- 1 sokolovm amazon 516M Sep 13 13:26 _0.cfs >>>>> > >> >>> -rw-r--r-- 1 sokolovm amazon 340 Sep 13 13:26 _0.si >>>>> > >> >>> >>>>> > >> >>> On Tue, Sep 13, 2022 at 8:50 AM Michael Sokolov < >>>>> msoko...@gmail.com> wrote: >>>>> > >> >>> > >>>>> > >> >>> > I ran another test. I thought I had increased the RAM >>>>> buffer size to >>>>> > >> >>> > 8G and heap to 16G. However I still see two segments in the >>>>> index that >>>>> > >> >>> > was created. And looking at the infostream I see: >>>>> > >> >>> > >>>>> > >> >>> > dir=MMapDirectory@ >>>>> /local/home/sokolovm/workspace/knn-perf/glove-100-angular.hdf5-train-200-200.index >>>>> > >> >>> > lockFactory=org\ >>>>> > >> >>> > .apache.lucene.store.NativeFSLockFactory@4466af20 >>>>> > >> >>> > index= >>>>> > >> >>> > version=9.4.0 >>>>> > >> >>> > >>>>> analyzer=org.apache.lucene.analysis.standard.StandardAnalyzer >>>>> > >> >>> > ramBufferSizeMB=8000.0 >>>>> > >> >>> > maxBufferedDocs=-1 >>>>> > >> >>> > ... >>>>> > >> >>> > perThreadHardLimitMB=1945 >>>>> > >> >>> > ... >>>>> > >> >>> > DWPT 0 [2022-09-13T02:42:53.329404950Z; main]: flush >>>>> postings as >>>>> > >> >>> > segment _6 numDocs=555373 >>>>> > >> >>> > IW 0 [2022-09-13T02:42:53.330671171Z; main]: 0 msec to >>>>> write norms >>>>> > >> >>> > IW 0 [2022-09-13T02:42:53.331113184Z; main]: 0 msec to >>>>> write docValues >>>>> > >> >>> > IW 0 [2022-09-13T02:42:53.331320146Z; main]: 0 msec to >>>>> write points >>>>> > >> >>> > IW 0 [2022-09-13T02:42:56.424195657Z; main]: 3092 msec to >>>>> write vectors >>>>> > >> >>> > IW 0 [2022-09-13T02:42:56.429239944Z; main]: 4 msec to >>>>> finish stored fields >>>>> > >> >>> > IW 0 [2022-09-13T02:42:56.429593512Z; main]: 0 msec to >>>>> write postings >>>>> > >> >>> > and finish vectors >>>>> > >> >>> > IW 0 [2022-09-13T02:42:56.430309031Z; main]: 0 msec to >>>>> write fieldInfos >>>>> > >> >>> > DWPT 0 [2022-09-13T02:42:56.431721622Z; main]: new segment >>>>> has 0 deleted docs >>>>> > >> >>> > DWPT 0 [2022-09-13T02:42:56.431921144Z; main]: new segment >>>>> has 0 >>>>> > >> >>> > soft-deleted docs >>>>> > >> >>> > DWPT 0 [2022-09-13T02:42:56.435738086Z; main]: new segment >>>>> has no >>>>> > >> >>> > vectors; no norms; no docValues; no prox; freqs >>>>> > >> >>> > DWPT 0 [2022-09-13T02:42:56.435952356Z; main]: >>>>> > >> >>> > flushedFiles=[_6_Lucene94HnswVectorsFormat_0.vec, _6.fdm, >>>>> _6.fdt, _6_\ >>>>> > >> >>> > Lucene94HnswVectorsFormat_0.vem, _6.fnm, _6.fdx, >>>>> > >> >>> > _6_Lucene94HnswVectorsFormat_0.vex] >>>>> > >> >>> > DWPT 0 [2022-09-13T02:42:56.436121861Z; main]: flushed >>>>> codec=Lucene94 >>>>> > >> >>> > DWPT 0 [2022-09-13T02:42:56.437691468Z; main]: flushed: >>>>> segment=_6 >>>>> > >> >>> > ramUsed=1,945.002 MB newFlushedSize=1,065.701 MB \ >>>>> > >> >>> > docs/MB=521.134 >>>>> > >> >>> > >>>>> > >> >>> > so I think it's this perThreadHardLimit that is triggering >>>>> the >>>>> > >> >>> > flushes? TBH this isn't something I had seen before; but >>>>> the docs say: >>>>> > >> >>> > >>>>> > >> >>> > /** >>>>> > >> >>> > * Expert: Sets the maximum memory consumption per thread >>>>> triggering >>>>> > >> >>> > a forced flush if exceeded. A >>>>> > >> >>> > * {@link DocumentsWriterPerThread} is forcefully flushed >>>>> once it >>>>> > >> >>> > exceeds this limit even if the >>>>> > >> >>> > * {@link #getRAMBufferSizeMB()} has not been exceeded. >>>>> This is a >>>>> > >> >>> > safety limit to prevent a {@link >>>>> > >> >>> > * DocumentsWriterPerThread} from address space >>>>> exhaustion due to >>>>> > >> >>> > its internal 32 bit signed >>>>> > >> >>> > * integer based memory addressing. The given value must >>>>> be less >>>>> > >> >>> > that 2GB (2048MB) >>>>> > >> >>> > * >>>>> > >> >>> > * @see #DEFAULT_RAM_PER_THREAD_HARD_LIMIT_MB >>>>> > >> >>> > */ >>>>> > >> >>> > >>>>> > >> >>> > On Mon, Sep 12, 2022 at 6:28 PM Michael Sokolov < >>>>> msoko...@gmail.com> wrote: >>>>> > >> >>> > > >>>>> > >> >>> > > Hi Mayya, thanks for persisting - I think we need to >>>>> wrestle this to >>>>> > >> >>> > > the ground for sure. In the test I ran, RAM buffer was >>>>> the default >>>>> > >> >>> > > checked in, which is weirdly: 1994MB. I did not >>>>> specifically set heap >>>>> > >> >>> > > size. I used maxConn/M=200. I'll try with larger buffer >>>>> to see if I >>>>> > >> >>> > > can get 9.4 to produce a single segment for the same test >>>>> settings. I >>>>> > >> >>> > > see you used a much smaller M (16), which should have >>>>> produced quite >>>>> > >> >>> > > small graphs, and I agree, should have been a single >>>>> segment. Were you >>>>> > >> >>> > > able to verify the number of segments? >>>>> > >> >>> > > >>>>> > >> >>> > > Agree that decrease in recall is not expected when more >>>>> segments are produced. >>>>> > >> >>> > > >>>>> > >> >>> > > On Mon, Sep 12, 2022 at 1:51 PM Mayya Sharipova >>>>> > >> >>> > > <mayya.sharip...@elastic.co.invalid> wrote: >>>>> > >> >>> > > > >>>>> > >> >>> > > > Hello Michael, >>>>> > >> >>> > > > Thanks for checking. >>>>> > >> >>> > > > Sorry for bringing this up again. >>>>> > >> >>> > > > First of all, I am ok with proceeding with the Lucene >>>>> 9.4 release and leaving the performance investigations for later. >>>>> > >> >>> > > > >>>>> > >> >>> > > > I am interested in what's the maxConn/M value you used >>>>> for your tests? What was the heap memory and the size of the RAM buffer >>>>> for >>>>> indexing? >>>>> > >> >>> > > > Usually, when we have multiple segments, recall should >>>>> increase, not decrease. But I agree that with multiple segments we can see >>>>> a big drop in QPS. >>>>> > >> >>> > > > >>>>> > >> >>> > > > Here is my investigation with detailed output of the >>>>> performance difference between 9.3 and 9.4 releases. In my tests I used a >>>>> large indexing buffer (2Gb) and large heap (5Gb) to end up with a single >>>>> segment for both 9.3 and 9.4 tests, but still see a big drop in QPS in >>>>> 9.4. >>>>> > >> >>> > > > >>>>> > >> >>> > > > Thank you. >>>>> > >> >>> > > > >>>>> > >> >>> > > > >>>>> > >> >>> > > > >>>>> > >> >>> > > > >>>>> > >> >>> > > > >>>>> > >> >>> > > > On Fri, Sep 9, 2022 at 12:21 PM Alan Woodward < >>>>> romseyg...@gmail.com> wrote: >>>>> > >> >>> > > >> >>>>> > >> >>> > > >> Done. Thanks! >>>>> > >> >>> > > >> >>>>> > >> >>> > > >> > On 9 Sep 2022, at 16:32, Michael Sokolov < >>>>> msoko...@gmail.com> wrote: >>>>> > >> >>> > > >> > >>>>> > >> >>> > > >> > Hi Alan - I checked out the interval queries patch; >>>>> seems pretty safe, >>>>> > >> >>> > > >> > please go ahead and port to 9.4. Thanks! >>>>> > >> >>> > > >> > >>>>> > >> >>> > > >> > Mike >>>>> > >> >>> > > >> > >>>>> > >> >>> > > >> > On Fri, Sep 9, 2022 at 10:41 AM Alan Woodward < >>>>> romseyg...@gmail.com> wrote: >>>>> > >> >>> > > >> >> >>>>> > >> >>> > > >> >> Hi Mike, >>>>> > >> >>> > > >> >> >>>>> > >> >>> > > >> >> I’ve opened >>>>> https://github.com/apache/lucene/pull/11760 as a small bug fix PR for >>>>> a problem with interval queries. Am I OK to port this to the 9.4 branch? >>>>> > >> >>> > > >> >> >>>>> > >> >>> > > >> >> Thanks, Alan >>>>> > >> >>> > > >> >> >>>>> > >> >>> > > >> >> On 2 Sep 2022, at 20:42, Michael Sokolov < >>>>> msoko...@gmail.com> wrote: >>>>> > >> >>> > > >> >> >>>>> > >> >>> > > >> >> NOTICE: >>>>> > >> >>> > > >> >> >>>>> > >> >>> > > >> >> Branch branch_9_4 has been cut and versions updated >>>>> to 9.5 on stable branch. >>>>> > >> >>> > > >> >> >>>>> > >> >>> > > >> >> Please observe the normal rules: >>>>> > >> >>> > > >> >> >>>>> > >> >>> > > >> >> * No new features may be committed to the branch. >>>>> > >> >>> > > >> >> * Documentation patches, build patches and serious >>>>> bug fixes may be >>>>> > >> >>> > > >> >> committed to the branch. However, you should submit >>>>> all patches you >>>>> > >> >>> > > >> >> want to commit to Jira first to give others the >>>>> chance to review >>>>> > >> >>> > > >> >> and possibly vote against the patch. Keep in mind >>>>> that it is our >>>>> > >> >>> > > >> >> main intention to keep the branch as stable as >>>>> possible. >>>>> > >> >>> > > >> >> * All patches that are intended for the branch >>>>> should first be committed >>>>> > >> >>> > > >> >> to the unstable branch, merged into the stable >>>>> branch, and then into >>>>> > >> >>> > > >> >> the current release branch. >>>>> > >> >>> > > >> >> * Normal unstable and stable branch development may >>>>> continue as usual. >>>>> > >> >>> > > >> >> However, if you plan to commit a big change to the >>>>> unstable branch >>>>> > >> >>> > > >> >> while the branch feature freeze is in effect, think >>>>> twice: can't the >>>>> > >> >>> > > >> >> addition wait a couple more days? Merges of bug >>>>> fixes into the branch >>>>> > >> >>> > > >> >> may become more difficult. >>>>> > >> >>> > > >> >> * Only Jira issues with Fix version 9.4 and >>>>> priority "Blocker" will delay >>>>> > >> >>> > > >> >> a release candidate build. >>>>> > >> >>> > > >> >> >>>>> > >> >>> > > >> >> >>>>> --------------------------------------------------------------------- >>>>> > >> >>> > > >> >> To unsubscribe, e-mail: >>>>> dev-unsubscr...@lucene.apache.org >>>>> > >> >>> > > >> >> For additional commands, e-mail: >>>>> dev-h...@lucene.apache.org >>>>> > >> >>> > > >> >> >>>>> > >> >>> > > >> >> >>>>> > >> >>> > > >> > >>>>> > >> >>> > > >> > >>>>> --------------------------------------------------------------------- >>>>> > >> >>> > > >> > To unsubscribe, e-mail: >>>>> dev-unsubscr...@lucene.apache.org >>>>> > >> >>> > > >> > For additional commands, e-mail: >>>>> dev-h...@lucene.apache.org >>>>> > >> >>> > > >> > >>>>> > >> >>> > > >> >>>>> > >> >>> > > >> >>>>> > >> >>> > > >> >>>>> --------------------------------------------------------------------- >>>>> > >> >>> > > >> To unsubscribe, e-mail: >>>>> dev-unsubscr...@lucene.apache.org >>>>> > >> >>> > > >> For additional commands, e-mail: >>>>> dev-h...@lucene.apache.org >>>>> > >> >>> > > >> >>>>> > >> >>> >>>>> > >> >>> >>>>> --------------------------------------------------------------------- >>>>> > >> >>> To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org >>>>> > >> >>> For additional commands, e-mail: dev-h...@lucene.apache.org >>>>> > >> >>> >>>>> > >> > >>>>> > >> > >>>>> > >> > -- >>>>> > >> > Adrien >>>>> > >> >>>>> > >> >>>>> --------------------------------------------------------------------- >>>>> > >> To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org >>>>> > >> For additional commands, e-mail: dev-h...@lucene.apache.org >>>>> > >> >>>>> > > >>>>> > > >>>>> > > -- >>>>> > > Adrien >>>>> >>>>> --------------------------------------------------------------------- >>>>> To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org >>>>> For additional commands, e-mail: dev-h...@lucene.apache.org >>>>> >>>>>
