[jdk17] RFR: 8271064: ZGC several jvm08 perf regressions after JDK-8268372

Tue Aug 3 08:58:31 UTC 2021

On Tue, 3 Aug 2021 08:22:58 GMT, Albert Mingkun Yang <ayang at openjdk.org> wrote:

>> After the integration of ZGC dynamic thread selection, there was a regression that popped up on a few benchmarks that all had in common a high allocation rate and very quick GC times. An extra conservative measure for not trusting the heuristics when using few GC threads, did not play very well with this scenario. The effect of it is back-to-back GC causing thousands of GCs when <100 were really needed. Removing this conservative measure causes the regression to go away, and a bunch of green benchmarks to pop up instead of red ones. This suggest that conservative measure isn't working very well and does not scale to these scenarios. It is also already taken care of by other conservative measures. I propose it is removed, to solve the regressions.
>> The entire ZGC perf suite was rerun (no red, only green), and tier1-3 was run. There isn't much potential for things to blow up as I just removed one parameter in the heuristic. So if anything nasty was to show up, it would be in the perf runs.
>
> Marked as reviewed by ayang (Committer).

Thanks for the reviews, @albertnetymk, @pliden and @tschatzl

-------------

PR: https://git.openjdk.java.net/jdk17/pull/298