On Fri, Jun 5, 2020 at 6:29 PM Mark Thomas <[email protected]> wrote: > On 05/06/2020 14:08, Martin Grigorov wrote: > > For load testing it I use wrk: wrk -c96 -t8 -d60s https://host:port/test > > The GC JVM arguments are: -Xloggc:./gc.log -XX:+PrintGCDetails > -verbose:gc > > JDK 1.8.0 b252 > > > > The GC logs are: > > - Tomcat 9.0.x (uses > > SynchronizedQueue/Stack): > https://gist.github.com/martin-g/d2570e7a6896e4d094ce548ceea3adb6 > > - Tomcat 9.0.x with my > > changes: > https://gist.github.com/martin-g/52c7d3a883b37e9bcd11ad6430800852 > > I've uploaded them to https://gceasy.io/ and the charts are similar. > > lock-free has marginally higher throughput. > > It has a longer GC pause but that might be a result of running longer > (~x2 longer). > > Average creation rate and average promotion rate are lower. > > Given this was with Java 8 I'm +1 to switching Tomcat 10 and also +1 to > switching Tomcat 9 if we can do it without impacting the API. >
I've noticed org.apache.tomcat.util.collections.TesterPerformanceSynchronizedStack and org.apache.tomcat.util.collections.TesterPerformanceSynchronizedQueue in test/ folder. Those give 1.5-3 times better results for SynchronizedQueue/Stack than Java Collections impls. I am not so sure my proposal is a good one anymore. > > I'm neutral on making the change in Tomcat 8.5 and -0.5 on Tomcat 7. > > Mark > > --------------------------------------------------------------------- > To unsubscribe, e-mail: [email protected] > For additional commands, e-mail: [email protected] > >
