RFR (XS): 8150952: Net PLAB size is clipped to max PLAB size as a whole, not on a per thread basis

Jesper Wilhelmsson jesper.wilhelmsson at oracle.com
Thu Mar 10 13:49:39 UTC 2016


Looks good!
/Jesper

Den 10/3/16 kl. 14:16, skrev Thomas Schatzl:
> Hi all,
>
>    could I have a "R"eviewer review for this change?
>
> Thanks,
>    Thomas
>
> On Thu, 2016-03-03 at 15:07 +0100, Thomas Schatzl wrote:
>> Hi all,
>>
>>    can I have reviews for this change that fixes PLAB sizing clipping
>> after we changed the desired PLAB size to be recorded without taking
>> the number of threads to use into account?
>>
>> The problem is that after calculating the net desired plab size, we
>> clip it to min/max sizes. This means that, when handing out the
>> desired
>> PLAB sizes depending on the number of threads, if the number of
>> threads
>> is really large, the PLAB sizes get way too small.
>>
>> I have seen some application logs where due to high number of
>> threads,
>> around 90%(!) of multiple GB of allocations during GC is direct
>> -allocated. Instead of using PLABs...
>>
>> This change fixes this issue.
>>
>> CR:
>> https://bugs.openjdk.java.net/browse/JDK-8150952
>> Webrev:
>> cr.openjdk.java.net/~tschatzl/8150952/webrev
>> Testing:
>> jprt, vm.gc, looking at several benchmarks to see whether the problem
>> still occurs
>>
>> Thanks,
>>    Thomas


More information about the hotspot-gc-dev mailing list