On Thu, 23 Mar 2023 08:20:18 GMT, Johannes Bechberger <jbechber...@openjdk.org> wrote:
>> Fixes the issue by transitioning the thread into the WXWrite mode while >> walking the stack in AsyncGetCallTrace. >> >> Tested on my M1 mac. > > Johannes Bechberger has updated the pull request incrementally with one > additional commit since the last revision: > > Use raw_thread instead of Thread::current() I tested the execution time of individual ASGCT calls on the renaissance benchmark using [asgct_perf_test](https://github.com/parttimenerd/asgct_perf_test) which I created for this purpose. TLDR: No relevant performance difference when disabling cache modification for ASGCT via a field in `Thread`. On Linux (64 core machine), all timings are in microseconds. With the disabled caches we get: bucket % count min mean max std std/mean median 90th 99th overall 100 81265409 0.04 4.07 59334.88 32.23 7.93 2.62 7.14 23.20 vs with current head OpenJDK: overall 100 81607301 0.03 3.96 83194.84 30.49 7.69 2.57 7.00 22.27 The difference of the averages just 110ns which should be undetectable in real life. With the disabled caches on Mac M1 we get: overall 100 39281484 0.00 1.39 94833.66 36.48 26.21 0.92 2.08 9.46 vs with the current JDK and [the fix adapted from async-profiler](https://github.com/async-profiler/async-profiler/blob/e3b7bfca227ae5c916f00abfacf0e61291df3a67/src/profiler.cpp#L383) which sets the WX mode: overall 100 37725188 0.00 1.40110634.45 29.39 20.92 0.92 2.17 10.21 Which doesn't alter the performance characteristics. I'm therefore pushing my new change to this PR, which prevents ASGCT from modifying any PcDesc cache. ------------- PR Comment: https://git.openjdk.org/jdk/pull/13144#issuecomment-1482570550
