viliam-durina commented on PR #15590: URL: https://github.com/apache/lucene/pull/15590#issuecomment-3777168935
The code fixed in this PR was introduced in 10.2.2 in #14527. We internally have been just recently upgrading from 10.2.1 to 10.2.2, and we noticed around 15% performance degradation of vector indexing. In our test we've been using 8 threads. Benchmarks were done as a part of the PR, but glancing through the discussion, it seems they were done before this particular code was introduced. With multiple threads, writes to `graphRamBytesUsed` are very frequent and thus very contended. This would be true if `AtomicLong` was used instead of `LongAdder`. The degradation would likely be much worse than 15% with more cores. After applying this PR to our internal 10.2.2 fork, we see the original performance. Because of the simplicity of the fix, I suggest backporting it to 10.3, and perhaps even to 10.2. -- 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]
