Can you boil this down to a microbenchmark (e.g. JMH) so you can look at assembly?
Maybe with on-heap, the dot product is getting vectorized, but with off-heap/unsafe it is not. e.g. something like this recent bug: https://bugs.openjdk.java.net/browse/JDK-8257531 You could re-run your bench on a JDK-16 early access as well, with that fix, and see what happens. On Wed, Dec 30, 2020 at 9:00 AM Michael Sokolov <msoko...@gmail.com> wrote: > > Hi, I've been working on improving performance of vector KNN search, > and found some behavior that seemed surprising to me, showing huge > differences in some cases comparing on-heap memory access with the way > we access data today via IndexInput. I'd love to get some other eyes > on this to help me better understand the difference, and whether it's > significant in reality or just an oddity that only turns up in a > microbenchmark. Thanks for taking a look (note: I don't plan to commit > the attached PR, it's just posted to show how the measurements were > done) > > https://github.com/apache/lucene-solr/pull/2173 > > --------------------------------------------------------------------- > 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
