Native Mark-Sweep: the more the warmups, the more the GCs

[chiarameiohas]

I noticed that in Mark-Sweep, the more the warmups, the more GCs are run. This does not happen on the same benchmark when running on Immix.

Mark-Sweep

• 1 warmup: 40 GCs
• 2 warmups: 41 GCs
• 3 warmups: 42 GCs
• 4 warmups: 52 GCs
• 5 warmups: 60 GCs

Immix

The first 4 all ran 40 GCs and the last one (5 warmups) ran 39 GCs.

Configuration

I used the mmtk-core version 440b57c with the OpenJDK binding.
I merged the commits from issue #824 (fixes a memory leak found in Mark-Sweep).
I used the biojava benchmark on a 500M heap.

This is the command I used:

MMTK_PLAN=MarkSweep  ./build/linux-x86_64-normal-server-release/images/jdk/bin/java -Djava.library.path=`pwd` -cp `pwd`:./dacapo-evaluation-git-29a657f.jar -XX:+UseThirdPartyHeap  -XX:MetaspaceSize=500M -Xms$HEAP_SIZE -Xmx$HEAP_SIZE Harness -c DacapoChopinCallback -n 3 biojava 

[k-sareen]

Probably related to the original bug you found in #824. The memory leak would mean that each iteration may do more GCs. One way to test if it's related/is a memory leak is to run minheap with 1, 5, and 10 iterations.

[chiarameiohas]

@k-sareen I forgot to mention that this was after I merged the changes from #824.
I'll run the checks with minheap and post the results here.