On Sat, 13 Sep 2025 00:19:23 GMT, Chris Plummer <cjplum...@openjdk.org> wrote:
> Improved test stability. See first comment for details. > > Tested by running the test 25 times with every GC on every supported platform > and also combined with -Xcomp. This test has proven very temperamental and unstable, so I've made a number of changes to address the stability issues. One realization while working on this was the ZGC youngGen using up the entire heap before a GC was forced, finally promoting to the oldgen and triggering the notification. This put the test at risk of unexpected OOMEs. G1 had a similar issue with delays getting objects promoted. The main way this has been addressed is by adding periodic System.gc() calls. It's not necessary for each call to force the promotion of all objects, but only that they get there over time as more calls are made. In order to avoid OOME due to the memory consumption loop running faster than the notification can be handled, the test now periodically queries the pool to see if more than 50% has been used. Once that happens it blocks until the notification has been received. Another tweak was to bump up the threshold to 20%, allowing for using a smaller heap (128m instead of 256m) without risk of exceeding the threshold before the notification is setup, which resulted in timeouts. The test had no checks to guarantee that the notification was being setup on the oldgen. For most GCs it got lucky and ended up with the right pool. However, for Shenandoah it was not using the right pool. In order to avoid this issue and detect errors with the notification setup, the notification is setup on pools with specific names so we are sure to get the right one. ------------- PR Comment: https://git.openjdk.org/jdk/pull/27266#issuecomment-3287233520
