Intermittent issue with long concurrent marking phase

Ramki Ramakrishna y.s.ramakrishna at oracle.com
Fri Sep 30 08:50:02 PDT 2011


Hi Srini --

It is true that the livelock can in fact resolve itself in some cases 
provided
there are external actions by mutators that result in that resolution.
So longish stalls that resolve themselves are indeed possible.
However, sometimes that action will never be taken by mutators because all
of them are already stalled, waiting for GC, so they can complete an 
allocation.

Thus, the lack of a complete stall does not rule out 6692906,
nor the presence of intermittent stalls confirm 6692906.

If your "stall" during the concurrent marking phase is easily detectable
by an external agent and you can take a pstack dump of the process in
that stalled phase, we may be able to identify with high confidence if 
this is
6692906 from a stack retrace of the process.

As regards experimenting with the workaround of turning off
CMSConcurrentMTEnabled.  It's possible that the overly long
marking cycle may have been a consequence of the bug, so turning
off MT-marking via -XX:-CMSConcurrentMTEnabled might well
be a workable solution for you without having to upgrade to get
the fix for 6692906.

It is of course easiest if you could reproduce this problem in a test 
set-up,
so you could do these experiments easily and make the right
determination.

-- ramki

On 9/30/2011 6:05 AM, Srini Padman wrote:
> Hi Ramki,
>   
> Apologies, both for mis-reading your original response (re: long initial-mark phases) and for choosing the wrong list. Thank you very much for redirecting it to gc-use.
>
> I just want to clarify a couple of points from your last response, for the record.
>
> To answer the question about the long stop-the-world initial marking phase: this is the longest I know of, but we have seen other instances where it lasted 3-4 seconds. In those cases as well, the "user/sys" times were much smaller than the "real" time so things clearly seem to be completely stalled. Also as a matter of background - the reason we moved to using the CMS collectors was that, prior to this, we were occasionally seeing extremely long (sometimes lasting more than a minute) full GCs. It is quite possible that the same factors that caused such long full GCs in the past are causing somewhat shorter (but still not _short_) initial mark with the CMS collector. In any case, I didn't view this as being related to 6692906 to begin with, and am glad to get confirmation that you don't think it is either.
>
> Regarding your point "A definite symptom of 6692906 can be diagnosed if the JVM completely stalls into a livelock (a few threads in the JVM intermittently active, but your application will forever stop  forward progress from that point on). It doesn't look like you have observed that latter symptom however?" - my understanding of the symptom (based on the summary at http://mail.openjdk.java.net/pipermail/hotspot-gc-dev/2010-February/001549.html, for example) was that the wait/stall resolves itself after tens of seconds for reasons unknown. Our symptom is closer to this, in that the application does not stop _forever_ from that point onwards. [It is entirely possible that there were further findings beyond what is in the posting that I am not aware of.]
>
> We will initiate efforts to reach out to JVM support - of course, in the meanwhile, any feedback or help on this forum is very welcome!
>
> Regards,
> Srini.
>
> --- On Thu, 9/29/11, Y. S. Ramakrishna<y.s.ramakrishna at oracle.com>  wrote:
>
>
> From: Y. S. Ramakrishna<y.s.ramakrishna at oracle.com>
> Subject: Re: Intermittent issue with long concurrent marking phase
> To: "Srini Padman"<srini_was at yahoo.com>
> Cc: hotspot-gc-use at openjdk.java.net
> Date: Thursday, September 29, 2011, 8:54 PM
>
>
> Hi Srini -- As I indicated, if you cannot upgrade easily to test
> if the issue is fixed, you should probably engage JVM support to
> get to a proper diagnosis of the issue affecting your production systems.
>
> more inline below ...
>
> On 09/29/11 07:44, Srini Padman wrote:
>> Hi Ramki,
>>    Thank you very much for your reply.
>>    It is not *always* that the concurrent marking phase takes this long, although it happens often enough. For example, in the full GC log corresponding to the snippet I pasted in my posting (attached, zipped) there is only that one instance.
> I see that there is one instance of the long _initial-mark_ pause
> and as i stated the whole process seems stalled at that time.
> It's definitely not the stall/livelock issue of CR 6692906.
>
> [In other words, you may be dealing with several different issues
> here and you will need to disentangle them.]
>
>>    I think I know why you are asking - based on my understanding of Bug # 6692906 (more accurately, based on discussions around it on this list), I was under the impression that such long CM phases will happen all the time (if they happen at all). Does the fact that it is intermittent raise the possibility that this is a different issue? I realize that you might not be able to answer this based on the bits of information you have, but perhaps the full GC log will tell you something that you don't already know.
> Again, my question about once/many was only about the long initial mark pause
> of which there is exactly one in your log.
>
> The stall of mutators during concurrent marking, which you are
> conjecturing is 6692906, is a different issue: for that you should
> either upgrade and test, or seek JVM support help. It is definitely
> possible that the symptoms of 6692906 will happen infrequently or
> intermittently. A definite symptom of 6692906 can be diagnosed if
> the JVM completely stalls into a livelock (a few threads in the
> JVM intermittently active, but your application will forever
> stop forward progress from that point on). It doesn't look like
> you have observed that latter symptom however?
>
> Sorry I can't really help more at this time, but perhaps someone
> from the community may be able to. But I really suggest either
> an upgrade or seeking JVM support help. (This is not a professional
> support alias.)
>
> In general, the GC-use alias is better suited to questions such as this.
> GC-dev should be used for GC development questions involving
> the main development trunk. Issues of uses in the field of
> older versions should be addressed to hotspot-gc-use at o.j.n
> So I've taken the liberty to send this to hotspot-gc-use at o.j.n
>
> All the best!
> -- ramki
>
>>    Regards,
>> Srini.
>>
>> --- On *Thu, 9/29/11, Ramki Ramakrishna /<y.s.ramakrishna at oracle.com>/* wrote:
>>
>>
>>       From: Ramki Ramakrishna<y.s.ramakrishna at oracle.com>
>>       Subject: Re: Intermittent issue with long concurrent marking phase
>>       To: "Srini Padman"<srini_was at yahoo.com>
>>       Cc: hotspot-gc-dev at openjdk.java.net
>>       Date: Thursday, September 29, 2011, 4:24 AM
>>
>>       Hi Srini -- (inline below)
>>
>>       On 9/28/2011 4:50 AM, Srini Padman wrote:
>>>       Questions:
>>>            1\ is it clear based on the description above that the issue is
>>>       identical to 6692906 (http://bugs.sun.com/view_bug.do?bug_id=6692906)?
>>>
>>       Very likely the same bug.
>>
>>>       2\ will we benefit by upgrading to a more recent JRE [1.6.0_26
>>>       being the one under consideration]?
>>>
>>       Definitely worth trying.
>>
>>>       3\ I have seen recommendations to use
>>>       "-XX:-CMSConcurrentMTEnabled" on some web forums - but I have
>>>       concerns about this; if we don't allow for concurrent marking to
>>>       use multiple threads, then isnt there a danger of marking
>>>       proceeding so slowly that we might end up running out of memory
>>>       i.e., garbage created much faster than it is collected]?
>>>
>>       Your concerns are very legitimate (especially given the length of
>>       the concurrent mark phase) and the number of cores you have.
>>
>>>            Any help is greatly appreciated. Please let me know if any
>>>       additional information is needed at all. I haven't attached the
>>>       full GC log (it caused problems with posting) but will gladly send
>>>       it directly to anybody who would like.
>>>
>>       The long initial mark pause is definitely concerning -- Does it show
>>       up regularly
>>       in the GC logs or is the snippet above an anomaly? Curisously, as
>>       the process time
>>       shows, the user and system time are both low but the elapsed time is
>>       very large.
>>       That looks like a total stall of the process, and I have no conjectures
>>       based on available data.
>>
>>       I suggest talking with your Java support folk if you reproduce this
>>       after upgrading to
>>       6u28 (or whatever).
>>
>>       best regards.
>>       -- ramki
>>
>>>       Regards,
>>>       Srini.
>>>
_______________________________________________
hotspot-gc-use mailing list
hotspot-gc-use at openjdk.java.net
http://mail.openjdk.java.net/mailman/listinfo/hotspot-gc-use


More information about the hotspot-gc-dev mailing list