RFR(XXS): 8153176: ParOldGC's ParallelTaskTerminator peeks wrong TaskQueueSet

[Mikael Gerdin]

Hi Richard,

On 2016-03-31 17:31, Reingruber, Richard wrote:
> Hi Mikael,
>
>    Mikael> As I mentioned in the bug I think the same problem affects parallel
>    Mikael> reference processing as well, see RefProcTaskExecutor::execute in
>    Mikael> pcTasks.cpp.
>
> Oh cool you saw this! Should I change this, too?

Yes, it's just as incorrect as the one you originally found.

>
>    Mikael> Another fun thing is that if there are only object arrays on any
>    Mikael> thread's work stealing queue then work stealing also breaks down,
>    Mikael> because the terminator doesn't check the _objarray_queues.
>
> Yes, I noticed this as well. If another thread is processing a large object
> array, then the elements will show up on its marking stack. This will let the
> thread spinning in offer_termination() return and continue in
> StealMarkingTask::do_it(), where it will try to steal from the
> _objarray_queues. I thought this might be sufficient.

Ah, right. That should save us then.

/Mikael

>
> Cheers, Richard.
>
> -----Original Message-----
> From: Mikael Gerdin [mailto:mikael.gerdin at oracle.com]
> Sent: Donnerstag, 31. März 2016 16:17
> To: Reingruber, Richard <richard.reingruber at sap.com>; hotspot-gc-dev at openjdk.java.net
> Subject: Re: RFR(XXS): 8153176: ParOldGC's ParallelTaskTerminator peeks wrong TaskQueueSet
>
> Hi Richard,
>
> On 2016-03-31 15:55, Reingruber, Richard wrote:
>> Hi,
>>
>> I would like to contribute a fix for a bug in parallel old gc. The bug can cause long gc pauses, because work stealing while marking does not work as ParallelTaskTerminator peeks the wrong task queue set.
>>
>> A microbenchmark that produces an object graph with large linear parts consistently shows 3x shorter gc pauses with the fix (see bug report). In a similar synthetic test the bug sporadically caused pauses 10x longer than average.
>>
>> Webrev: http://www.sapjvm.com/rr/webrevs/8153176_paroldgc_wrong_taskqueueset_in_marking/webrev_01/
>
> As I mentioned in the bug I think the same problem affects parallel
> reference processing as well, see RefProcTaskExecutor::execute in
> pcTasks.cpp.
>
> Another fun thing is that if there are only object arrays on any
> thread's work stealing queue then work stealing also breaks down,
> because the terminator doesn't check the _objarray_queues.
>
> Perhaps as a future change ParallelTaskTerminator could be decoupled
> from TaskQueueSetSuper and instead using either template based
> polymorphism or virtual calls to do the "peek in queue set" callback.
>
> /Mikael
>
>> Bug:    https://bugs.openjdk.java.net/browse/JDK-8153176
>>
>> The change needs to be sponsored as well, please.
>>
>> Thanks, Richard.
>>
>> --
>> Richard Reingruber | SAP JVM | CORE PLATFORM,  SAP SE
>> Pflichtangaben/Mandatory Disclosure Statements: http://www.sap.com/company/legal/impressum.epx
>>