I think the nightly benchmarks may need to get beefed up with some compiler output so we can better see what is happening? openjdk supports many flags to print the compilation/inlining details.
On Fri, Aug 15, 2025 at 3:08 PM Adrien Grand <jpou...@gmail.com> wrote: > > Nightly benchmarks caught up with this change, and got significant slowdowns > as expected: https://benchmarks.mikemccandless.com/2025.08.13.20.38.43.html. > > It'd be good to better understand why it hurts so much and how it could be > fixed, help is welcome. :) > > On Wed, Aug 13, 2025 at 8:39 PM Adrien Grand <jpou...@gmail.com> wrote: >> >> Hello all, >> >> For reference, I just merged this PR[0] to luceneutil, which "pollutes" call >> sites of DocIdSetIterator#nextDoc, DocIdSetIterator#advance, Bits#get (live >> docs), SimScorer#score and others to better reflect what performance may >> look like on a production system that handles diverse workloads. >> >> A consequence is that luceneutil now report worse performance. In >> particular, nightly benchmarks are expected to get noticeable drops on many >> tasks. This behavior in enabled by default but it is possible to disable it >> by passing pollute=False when initializing a Competitor. >> >> [0] https://github.com/mikemccand/luceneutil/pull/436 >> >> -- >> Adrien > > > > -- > Adrien --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
