I did a quick run with your patch, but since I turned on the CMS as well as TieredMergePolicy I'm not sure how fair the comparison is. Here's the result: Candidate: Indexer: indexing done (890209 msec); total 33332620 docs Indexer: waitForMerges done (71622 msec) Indexer: finished (961877 msec) Baseline: Indexer: indexing done (909706 msec); total 33332620 docs Indexer: waitForMerges done (54775 msec) Indexer: finished (964528 msec)
For more accurate comparison I guess it's better to use LogxxMergePolicy and turn off CMS? If you want to run it yourself you can find the lines I quoted from the log file. Patrick On Mon, Apr 24, 2023 at 12:34 PM Thomas Dullien <thomas.dull...@elastic.co.invalid> wrote: > Hey all, > > I've been experimenting with fixing some low-hanging performance fruit in > the ElasticSearch codebase, and came across the fact that the MurmurHash > implementation that is used by ByteRef.hashCode() is reading 4 bytes per > loop iteration (which is likely an artifact from 32-bit architectures, > which are ever-less-important). I made a small fix to change the > implementation to read 8 bytes per loop iteration; I expected a very small > impact (2-3% CPU or so over an indexing run in ElasticSearch), but got a > pretty nontrivial throughput improvement over a few indexing benchmarks. > > I tried running Lucene-only benchmarks, and succeeded in running the > example from https://github.com/mikemccand/luceneutil - but I couldn't > figure out how to run indexing benchmarks and how to interpret the results. > > Could someone help me in running the benchmarks for the attached patch? > > Cheers, > Thomas > > --------------------------------------------------------------------- > To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org > For additional commands, e-mail: dev-h...@lucene.apache.org