RFR (XXS): 7110173: GCNotifier::pushNotification publishes stale data.

[Frederic Parain]

On 11/17/11 3:09 PM, Tony Printezis wrote:
> Fred,
>
> I haven't been following this thread closely, but I do have a question
> (see below).
>
> On 11/17/2011 8:43 AM, Frederic Parain wrote:
>> Hi Mandy,
>>
>> It would have been nice to have notifications providing
>> detailed information about CMS different phases, but it
>> doesn't seem easy to do right now. I did the experiments
>> with CMS using JConsole, and the result is not really
>> satisfactory:
>> - several notifications are sent for each CMS cycle
>> but the notification lacks the information to identify
>> the CMS phase associated with the notification (worse
>> than that, they all claim they notify the "end of a major
>> GC cycle" which is certainly not true for CMS.
>
> Are you referring to notifications at the end of a concurrent cycle
> (i.e., at the end of sweeping)? Depends on how you look at it, but a
> concurrent cycle is a major cycle given that it collects the entire
> heap. In fact, IIRC, the a concurrent cycle should increase the full gc
> count.

I'm not referring to the concurrent cycles, but to the different phases
of CMS. It seems that CMS is calling GCMemoryManager::gc_end() after
the initial marking, the final marking and the sweeping phase. The
gc count (_num_collections) is incremented only after the completion of
the sweeping phase. The actual notification code is confused
by the fact that CMS is executing GCMemoryManager::gc_end() several
times during a GC cycle. I made the assumption that calling gc_end() 
means that the GC cycle is completed, but this assumption is wrong with 
CMS, so the notification message is wrong in this case.

Fred

-- 
Frederic Parain - Oracle
Grenoble Engineering Center - France
Phone: +33 4 76 18 81 17
Email: Frederic.Parain at oracle.com