RFR(M) Enable UseDynamicNumberOfGCThreads flag support in Shenandoah

[Zhengyu Gu]

Last time I tried to kill one WorkGang, I ran into some problems which I can not recall at this moment.

I will try again.

Thanks,

-Zhengyu



On 02/06/2017 04:34 PM, Roman Kennke wrote:
> Sorry, this went under the radar with FOSDEM and all.
>
> The patch looks ok.
>
> We should think again about killing conc-workers and simply use a
> single dynamically-scaled thread pool.
>
> Roman
>
> Am Montag, den 06.02.2017, 16:27 -0500 schrieb Zhengyu Gu:
>> Resend with proper subject line.
>>
>> Thanks,
>>
>> -Zhengyu
>>
>>
>>
>> On 02/02/2017 07:39 AM, Zhengyu Gu wrote:
>>> The heuristics for selecting number of workers probably still need
>>> to
>>> be tweaked, as I did not see
>>> improvement with flag on/off with SPECjbb2015.
>>>
>>> The patch fixes a couple of potential issues:
>>>
>>> - Without the patch, enable UseDynamicNumberOfGCThreads can crash
>>> VM
>>> due to uninitialized gclab of
>>>    dynamic created workers.
>>>
>>> - Imbalance of task queues and workers that potential cause misses
>>> of
>>> final mark. The problem is hidden
>>>    because of currently implementation utilizing max parallel
>>> workers
>>> for the final mark from root.
>>>
>>> I would like to get this push if there are not problems with the
>>> patch
>>> itself, so we can find more
>>> systems to tweak the algorithm.
>>>
>>> Webrev:
>>> http://cr.openjdk.java.net/~zgu/shenandoah/dynamic_workers/webrev.0
>>> 0/
>>>
>>> Tests:
>>>    - SPECjbb2015 fastdebug and release
>>>    - SPECjvm     Debry fastdebug and release
>>>
>>> Thanks,
>>>
>>> -Zhengyu
>>>
>>