RFR(S): 7068240: G1: Long "parallel other time" and "ext root scanning" when running specific benchmark
jon.masamitsu at oracle.com
Fri Jul 29 17:48:21 UTC 2011
Looks good. Ship it.
On 7/29/2011 10:01 AM, John Cuthbertson wrote:
> Hi Everyone,
> I'm still looking for another review of this webrev.
> On 07/20/11 15:45, John Cuthbertson wrote:
>> Hi Everyone,
>> Can I have a couple of volunteers look over these changes? The webrev
>> can be found at: http://cr.openjdk.java.net/~johnc/7068240/webrev.0/
>> When running a specific benchmark, it was noticed that the "ext root
>> scanning" time was significantly longer for one thread than the
>> others - in a good number of the evacuation pauses. Some of these
>> long times were caused by the stack depth on an application thread
>> but others were caused by the scanning of the reference processor's
>> discovered lists. Furthermore the scanning of the discovered lists
>> was taking place after RSet updating and scanning, negating any
>> stealing that might take place in these parts, and causing longer
>> termination times. Also the scan closures were applied directly to
>> the discovered lists rather than a BufferingOopClosure being applied.
>> As a result, if any copying took place, the time for that copying was
>> being attributed to ext. root scanning rather than "object copying".
>> Additionally I also removed a couple statistic variables from
>> G1CollectorPolicy that were not thread-local but being updated and
>> (overwritten) by the parallel worker threads. The only place where
>> these statistic variables were being used was in an instrumentation
>> block that was executed when G1PolicyVerbose was set appropriately.
>> Testing: gc test suite and jprt.
More information about the hotspot-gc-dev