At a G1 performance meeting with the some SPARC hardware engineers, the SPARC engineers showed some G1 GC logs where the time attibuted to "GC Worker Other" time was much higher than expected (around 50ms). I believe the benchmark being run was SPECjbb2012 - but I don't have details of the VM flags they were running with. The system they were running on was, I believe, a T4. The SPARC engineers offered some collect experiements from the affected runs - which should help us track down this particular instance of the issue. We have seen this issue before and the "Other" time was coming from the freeing Chunks associated with the resource allocations. The resource allocations were coming resource-allocating a closure while scanning a card during RSet scanning. We were resource-allocating a new closure per card scanned.
|