Running some micro benchmarks with lock tracing in the Oracle Studio analyzer show a significant amount of time spent in stack traces (i.e. actual cpu time plus much sync wait time) with G1Allocator::par_allocate_during_gc() holding a lock. This amounts to around 5-10% of evacuation, and almost all sync wait time.

This is most likely the FreeList_lock grabbed before the call to G1AllocRegion::attempt_allocation_locked().

Investigate, and try to improve.