Understanding High Object Copy Times

Mon May 1 21:19:28 UTC 2017

Hi Kirk,

A comment from me embedded below.

charlie

> On May 1, 2017, at 3:42 PM, Kirk Pepperdine <kirk at kodewerk.com> wrote:
> 
> So, never mind all this.. I scrubbed all of the noise out of the log and what I see here is a Heap After GC chart that shows heap completely filled. Unless the Full GC cleaned things up, this VM isn’t going any further. Heap fills very quickly.
> 
> 
>> 
> 
>> 
>> Remember set is the data structure that keep track of pointers from other regions to the region we are operating on. If there are pointers from old regions to the objects in the young regions, we can not collect that object.
>> After a gc pause, the live objects are evacuated to either the survivor or old gen regions. Then we need to update the Remember set. That is what I was referring to.
>> It is part of object copy.
>> 
>> Like Kirk mentioned, the parallelism of the gc threads seems fine. Charlie.Hunt suggested maybe the object graph is very deep so that the gc threads can not steal work and end up spinning.
> 
> I’d like to hear more from Charlie on this point because if the parallelism is ok then I would conclude that the work load is well balanced implying that there shouldn’t be any work stealing.. am prepared to be wrong on this but…. and, that said the number of attempts to shutdown suggest that there was something going on at the tail end of the collection and maybe work stealing was far to granular???? No idea….

I haven’t had a chance to look at the logs in detail.  That said, generally if we see high termination times it suggests that one or more of the GC threads are in a termination protocol where it is waiting for one or more other GC threads to finish their work. Hence my thought that there may be some part of the object graph that is really deep that is keeping one GC thread very occupied, i.e. much longer than the other GC threads.

Btw, thanks for jumping in and offering a detailed analysis … a full heap followed by a Full GC is not good, especially if it didn’t make much space available. :-|

Ooh, before I forget, did I see Nezih mention that THP was enabled?  If that’s the case, I would disable it.

thanks,

charlie

> 
> Kind regards,
> Kirk
> 
> 
>> 
>> Some things we can try:
>> -XX:MaxGCPauseMillis=<200> Maybe increase it to 400
>> Maybe try to reduce the -XX:ParallelGCThreads
>> 
>> -XX:+G1SummarizeRSetStats
>> -XX:G1SummarizeRSetStatsPeriod=10
>> 
>> It is very expensive to print those statistics, for now we know there is a lot of coarsening. We can make G1SumarizeRSetStatsPeriod=100
>> 
>> The RSet footprint is so big,  increasing -XX:G1RSetRegionEntries will get rid of coarsening but make the memory footprint bigger. We can delay this for now.
>> 
>> What is your hardware/software configuration? cpu/memory/cores? There is no swapping, right?
>> 
>> Thanks
>> Jenny
>> 
>> On 04/27/2017 08:52 PM, nezih yigitbasi wrote:
>>> My first question is incorrect actually, so I am giving a better example to rephrase my question. At time "2017-04-26T03:12:24.259-0700"  there is a young GC that took 35.47 s, where object copy took 28983.4 ms. In that event I see the following log:
>>> 
>>>    [Eden: 7168.0M(7168.0M)->0.0B(7168.0M) Survivors: 1024.0M->1024.0M Heap: 153.4G(160.0G)->149.4G(160.0G)]
>>> 
>>> My interpretation for this is, ~4GB of garbage was collected from heap in total and we see that eden usage goes down by ~7GB, this means ~3GB of the eden was live objects. Is this interpretation correct? If it is, how come copying over 3GB takes ~29s? In your answer you said "most of the object copy time is dealing with the Remember Set", can you please give some details about what kind of operations on rsets are done during the object copy phase, and can we see that from these logs?
>>> 
>>> Thanks again,
>>> Nezih
>>> 
>>> 
>>> 2017-04-27 17:20 GMT-07:00 nezih yigitbasi <nezihyigitbasi at gmail.com <mailto:nezihyigitbasi at gmail.com>>:
>>> Thanks for the suggestions. We use the default pause time. And here is our entire set of JVM args: https://gist.github.com/nezihyigitbasi/04f5fdb9c32ac56097011819e20602d8 <https://gist.github.com/nezihyigitbasi/04f5fdb9c32ac56097011819e20602d8>
>>> 
>>> I have some followup questions:
>>> 
>>> - In some case the object copy took 39406.8 ms, even if the remembered set is ~30GB isn't this too slow (that's <1GB/s of data)?
>>> - Is there any way to reduce the rset overhead?
>>> - My initial thought when I saw the high object copy times was there may be some sort of contention to have such a low throughput during the copy. Although it may not be the case here, I just wonder whether there is a way to see the amount of contention from the gc logs?
>>> 
>>> Nezih
>>> 
>>> 
>>> 2017-04-27 16:58 GMT-07:00 Jenny Zhang <yu.zhang at oracle.com <mailto:yu.zhang at oracle.com>>:
>>> Hi, Hezih,
>>> 
>>> It seems this workload is very heavy on Remember Set. It has about 31G native memory for RSet (old gen) and still with coarsening.
>>> 
>>> What is you pause time goal? The default (200ms) might be too small for you.  Can you increase that so G1 can increase the young gen size? Since there is not much promotion, I guess most of the object copy time is dealing with the Remember Set.
>>> 
>>> There are other things you can try, like increase the G1RSetReginEntries, but the memory footprint will be bigger.
>>> 
>>> So if you can, I suggest increase the pause time goal first.
>>> 
>>> Thanks
>>> 
>>> Jenny
>>> 
>>> 
>>> 
>>> 
>>> On 4/27/2017 9:22 AM, nezih yigitbasi wrote:
>>> Hi,
>>> We see huge object copy times (and relatively high termination times) during young GCs in our production system running on Java 1.8.0_112-b15. You can find the GC logs here: https://gist.github.com/nezihyigitbasi/1f7a92da7860908a611cb1197bd8626b <https://gist.github.com/nezihyigitbasi/1f7a92da7860908a611cb1197bd8626b>
>>> 
>>> The young GC times start going high after the timestamp "2017-04-26T03:07:22.164-0700".
>>> 
>>> I will appreciate if you can give some details about:
>>> - what goes into the "Object Copy" phase during young GCs and how we can reduce it.
>>> - why we see high Termination times and what we can do about it
>>> 
>>> Thanks,
>>> Nezih
>>> 
>>> 
>>> 
>> 
>> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/hotspot-gc-dev/attachments/20170501/f235ab37/attachment.htm>