Community
Participate
Working Groups
Performance tests run with default values for max heap size (which is currently set to 256Mb). It seems that performance tests run close to this limit (200 - 250Mb on my computer). This alters results as: - more time is spend in garbage collections; - components that use soft references might discard their contents. For instance, extension registry cache will be freed as we get low on memory resulting in substantial slowdown in all operations that use extension registry. If the heap size is increased to 512Mb, tests take 300 - 350Mb on my computer. I'd recommend increasing maximum heap size for the performance tests to 512Mb. (Not sure if this needs to be done separately for the test framework and for the inside-IDE runs.)
Can we start by capturing how many times the garbage collector runs during each test (and how much time this took)?
Created attachment 143726 [details] patch
Note that this must be done both in HEAD and baseline stream (perf_35x). I also wonder if we should not need also increase the Xms value to 512M. I'm afraid that if we let to 256M, then we got some extra time consumed by the VM when it needs some more room for memory and then makes the concerned test time having some strange behavior...
I've released this change to the perf_35x branch as well.
Fixed.