G1 STW phases and FGC column in jstat

[Yasumasa Suenaga]

Hi Kirk,

Agree to you.
However, I use occasional FGC counter at CMS as below:

  1. Check major collection occurrence
       Some production systems have large memory as Java heap,
       and they are not set GC log.
       If their CPU usage becomes high, I want to check GC execution.
       (Of course, we have to check any other points :-) )

  2. Core image analysis
       If JVM is crashed, I want to check PerfCounter to know situation.
       (In the past, I sometimes encountered crash at GC worker thread.)

I guess that I will want to check them at G1.

Thus, at least, I want to add PerfCounter for CGC (and add JVMTI event hook).
However, this proposal will affect to jstat spec.
So I want to discuss about it before filing to JBS.


Thanks,

Yasumasa


On 2016/03/02 19:02, kirk at kodewerk.com wrote:
> Hi Yasumasa,
>
> Good question. I’ve never considered CMS to be a Full GC. This implies that there should be separate performance counters for CMS pause phases as it is possible to have FGC. Of course a FGC during CMS maybe user triggered, triggered outside a CMS cycle, interrupts a CMS cycle, or interrupts a CMS phase. I’m not sure how much of a distinction one needs to make here as that could be a quick broader discussion. Certainly the purpose isn’t to recreate the GC logs in these performance counters. But at the very least not having a distinction between full and a STW-CMS phase is kind of misleading in my opinion.
>
> Regards,
> Kirk
>
>> On Mar 1, 2016, at 5:03 PM, Yasumasa Suenaga <yasuenag at gmail.com> wrote:
>>
>> Hi Kirk,
>>
>>> It is also incorrect to count initial mark and remark in CMS as a FGC.
>>
>> Though, how can we check execution of major collection without GC log?
>> Should we add new PerfCounter for CGC (and add CGC column to jstat output)?
>>
>>
>> Yasumasa
>>
>>
>> On 2016/03/02 6:35, kirk at kodewerk.com wrote:
>>> Hi,
>>>
>>> I think it is incorrect to count remark and cleanup as FGC. They are not full collections. It is also incorrect to count initial mark and remark in CMS as a FGC. It is unfortunate that this is counted this way.
>>>
>>> Regards,
>>> Kirk
>>>
>>>> On Mar 1, 2016, at 8:56 AM, Yasumasa Suenaga <yasuenag at gmail.com> wrote:
>>>>
>>>> Hi all,
>>>>
>>>> I wonder that STW phases (Remark and Cleanup) at G1 are not counted in jstat FGC column.
>>>> For example, Initial Mark and Remark at CMS are counted as FGC.
>>>>
>>>> For consistency, I think that G1 STW phases should be counted as FGC.
>>>> What do you think about it?
>>>>
>>>> If it is accepted, I will file it to JBS and will upload webrev.
>>>>
>>>>
>>>> suggested fix:
>>>> ----------------------
>>>> diff -r 8a103ba9a7b2 src/share/vm/gc/g1/g1MonitoringSupport.cpp
>>>> --- a/src/share/vm/gc/g1/g1MonitoringSupport.cpp	Mon Feb 29 22:54:24 2016 +0900
>>>> +++ b/src/share/vm/gc/g1/g1MonitoringSupport.cpp	Tue Mar 01 23:43:30 2016 +0900
>>>> @@ -103,7 +103,7 @@
>>>>    //   name "collector.1".  In a generational collector this would be the
>>>>    // old generation collection.
>>>>    _full_collection_counters =
>>>> -    new CollectorCounters("G1 stop-the-world full collections", 1);
>>>> +    new CollectorCounters("G1 stop-the-world phase", 1);
>>>>
>>>>    // timer sampling for all counters supporting sampling only update the
>>>>    // used value.  See the take_sample() method.  G1 requires both used and
>>>> diff -r 8a103ba9a7b2 src/share/vm/gc/g1/vm_operations_g1.cpp
>>>> --- a/src/share/vm/gc/g1/vm_operations_g1.cpp	Mon Feb 29 22:54:24 2016 +0900
>>>> +++ b/src/share/vm/gc/g1/vm_operations_g1.cpp	Tue Mar 01 23:43:30 2016 +0900
>>>> @@ -230,6 +230,8 @@
>>>>    G1CollectedHeap* g1h = G1CollectedHeap::heap();
>>>>    GCTraceTime(Info, gc) t(_printGCMessage, g1h->gc_timer_cm(), GCCause::_no_gc, true);
>>>>    IsGCActiveMark x;
>>>> +  SvcGCMarker sgcm(SvcGCMarker::OTHER);
>>>> +  TraceCollectorStats tcs(g1h->g1mm()->full_collection_counters());
>>>>    _cl->do_void();
>>>> }
>>>>
>>>> ----------------------
>>>>
>>>>
>>>> Thanks,
>>>>
>>>> Yasumasa
>>>>
>>>>
>>>
>