benwtrent commented on issue #11830: URL: https://github.com/apache/lucene/issues/11830#issuecomment-1279207529
OK, I ran this patch against minst, sift, glove, and deep image. Recall is the exact same in all cases, so no mistakes there. There are indeed slight differences in latency, below are images. I ran search 5 times for each parameter combination and chose the fastest. All of these were with `"M": 16, "efConstruction": 100`. For each, I specifically point out the QPS at the lowest recall. The deltas in QPS vary from near 0% change to 11% decrease. 11% does seem large but I think its an outlier. The median decrease is 2%, the average change is 0% (or 0.002%). [ann_qps_comparison - ann_scores.csv](https://github.com/apache/lucene/files/9788126/ann_qps_comparison.-.ann_scores.csv) For every data set, the memory required reduced at around a similar amount. This change seems like a good step forward to me. @jtibshirani # Deep Image  # Glove  # Minst  # Sift  -- 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: issues-unsubscr...@lucene.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
