During performance runs of SPECjEnterprise, a few extremely long remark pasues were observed ( > 5s). These long pauses caused the app server under test to prematurely end sessions/close connections resulting in a significant reduction of injection rate.

After analysis of the GC log and code it was that the reason for these long remark pauses might be the reference processing that takes place during a remark. By default the reference processing that takes place during a remark pause is single threaded (executed by the VMThread). Parallel reference processing is turned on by setting the ParallelRefProcEnabled flag to true and the number of ParallelGCThreads > 1.

Enabling parallel reference processing does incur an overhead - which we don't want to incur. Thererefore we want to be able to make the enabling of parallel reference processing dynamic, independent of the ParallelRefProcEnabled flag, and based upon the work that the reference processor has to do.