RamakrishnaChilaka commented on PR #15140: URL: https://github.com/apache/lucene/pull/15140#issuecomment-3253461571
> I'm looking forward to seeing whether this can help with vector search as well. cc @benwtrent @msokolov For this benchmark, I switched `FloatHeap` and `BlockingFloatHeap` from binary to ternary heaps, and used `TernaryLongHeap` in `NeighborQueue`. Luceneutil has been run with TopN as 1000 (see https://github.com/mikemccand/luceneutil/pull/357). I believe most of the speedup comes from the heap exchange in `MultiLeafKnnCollector`. Please find the results below (source: `enwiki-20120502-mpnet` (default in vector-test), Ran benchmarks on c8g.8xlarge (graviton) instance.) ``` TaskQPS baseline StdDevQPS candidate StdDev Pct diff p-value PKLookup 469.82 (0.8%) 473.96 (1.0%) 0.9% ( 0% - 2%) 0.002 AndHighMedVector 122.09 (1.8%) 124.39 (1.4%) 1.9% ( -1% - 5%) 0.000 AndHighLowVector 139.25 (2.0%) 142.01 (1.5%) 2.0% ( -1% - 5%) 0.000 AndHighHighVector 122.54 (1.8%) 124.97 (1.4%) 2.0% ( -1% - 5%) 0.000 HighTermVector 131.17 (1.8%) 133.95 (1.4%) 2.1% ( -1% - 5%) 0.000 MedTermVector 147.68 (1.9%) 150.89 (1.4%) 2.2% ( -1% - 5%) 0.000 LowTermVector 144.38 (2.0%) 147.60 (1.4%) 2.2% ( -1% - 5%) 0.000 ``` -- 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. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
