CI shows a strange 12% slowdown for JikesRVM/SemiSpace/lusearch

[qinsoon]

Starting from this PR #324, we sometimes observed our CI shows a 12% slowdown when comparing a PR with master for JikesRVM/SemiSpace/lusearch. We also observed this slowdown in a few PRs after that (for example: #334 (comment)). This slowdown randomly happens (sometimes we do not see the slowdown). If we see the slowdown, it always shows ~12% slowdown for a PR comparing with master for that one benchmark. In such case, the result for the master is strangely faster (~580ms) which normally should be around 610-620ms.

https://www.mmtk.io/ci-perf-result/master/jikesrvm_semispace_history.html shows that from the run 2021-04-19-Mon-161118, our results for lusearch starts to have a large variance that matches the slowdown we observed. We should investigate further into this.
The PRs that were merged before the run 2021-04-19-Mon-161118 are:

• Fix stress gc triggering in GenCopy Fix stress gc triggering in GenCopy #292
• Remove cyclic reference between space and page resource Remove cyclic reference between space and page resource (merge after PR #268) #275
• Fix race condition in FragmentedMmapper.ensure_mapped() Fix race condition in FragmentedMmapper.ensure_mapped() #278

We should investigate further into this.