https://bz.apache.org/bugzilla/show_bug.cgi?id=70305
--- Comment #2 from [email protected] --- Thanks for looking! To clarify, no, it isn't a leak. You're right that nothing is retained after GC, the problem is turnover. What we observed was under steady traffic a constant stream of classes being defined and unloaded. Engines that receive fewer than 127 callbacks never trigger it, which may be why your load tests didn't show it. When profiling two identical servers under the same load (FFM OpenSSL vs JSSE) we observed the following: Classes loaded / unloaded 464K / 428K vs 39K / 3K Concurrent mark (class unloading) 4.94 s per cycle vs 0.44 s GC in 60 s JFR window 4 degenerated + 1 full, 5.04s pause vs all concurrent, 50 ms total The servers have been up for a while, which explains the Classes loaded / uploaded numbers. The dead classes lengthen every Shenandoah mark cycle until it falls behind allocation, which cause slower GC. -- You are receiving this mail because: You are the assignee for the bug. --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
