This is part of the investigation of https://bugs.openjdk.java.net/browse/JDK-8034242.
bugdb bug https://bug.oraclecorp.com/pls/bug/webbug_edit.edit_info_top?rptno=18096739

5% of the regression is due to rt check in starting in b108.
From GC(CMS) logs, the survivors is 40M in jdk8b108 and 2M in jdk8b106.
If we run with b106 rt.jar in jdk8b108, the regression is gone.
From Heapdump, there are about 18M java.util.HashMap$Node and java.util.HashMap$Node[] which are not reachable.

I have instrumented HashMap.java put()/putVal methods to see if jdk8b108 treeify more nodes. Provided the rt.jar to PSR to run.

Intel is seeing increased survivors for the SPECjbb2013 run, which might be the same cause.