divijvaidya commented on PR #12890: URL: https://github.com/apache/kafka/pull/12890#issuecomment-1352933410
Hey @ijuma - thank you for your question. 1. I used [Amazon CodeGuru Profiler](https://docs.aws.amazon.com/codeguru/latest/profiler-ug/what-is-codeguru-profiler.html). I ran the test again with with [async-profiler](https://github.com/jvm-profiling-tools/async-profiler) which should not suffer from safepoint bias. Please find the below image for the usage of the function. It demonstrates that HashCode evaluation is not the major component but calculating the size takes quite some time. After this change, the flamegraph does not have this stack frame at all (since we don't parse the header again). The flamegraphs is in html format so I can't add it here but will be happy to share with you offline in case you are interested. I am running this with Correto JDK 11 and `./profiler.sh -d 500 -i 40ms -e cycles` as the profiler configuration.  2. Irrespective of the CPU usage optimisation that this change brings, I think that this is a good change because it avoids parsing the request header twice. 3. The actual CPU standard deviation is more than 2-3% during the duration of the test and hence, I cannot reliably say that I observed lower CPU usage after this change. Latency measurement is also not reliable because the bottleneck is thread contention on partition lock in UnifiedLog and hence, this change won't change the latency. -- 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: jira-unsubscr...@kafka.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org
