G1 discovers same garbage again?

Srinivas Ramakrishna ysr1729 at gmail.com
Mon Dec 12 19:40:54 UTC 2011


As I said, instead of relying on jconsole's asynchronous heap size samples,
look at the GC log to see
if the inter-scavenge period suddenly falls (i.e. the application is
presumably allocating at a much faster rate
than is normal). If the GC log confirns that, then see if turning off  TLAB
resizing (and fixing it to a reasonable
value) makes any difference.

Your existing GC logs should be sufficient to answer the first part of the
question above.

best.
-- ramki

On Mon, Dec 12, 2011 at 11:08 AM, Florian Binder <java at java4.info> wrote:

>  Tony,
>
> As you have seen at the jconsole-screenshot the app does make very few
> garbage. It is optimized to make no garbage. But suddenly there seems to be
> every 90 seconds about 4 gb of (new) garbage (even the load of the app does
> not change). Therefore I thought it might be the g1-collector. I will try
> to reproduce this senario an have a closure look on it.
>
> Flo
>
> Am 12.12.2011 15:46, schrieb Tony Printezis:
>
> Florian,
>
> Not sure what you mean by "the same GC". It is because the space
> transition looks like: 16394M->12914M(20480M)? This is normal. Heap
> occupancy reaches 16394M, GC happens, heap occupancy goes down to 12914M.
> Next time it happens, the heap will have new objects. This is how GCs work.
>
> Tony
>
> On 12/09/2011 09:08 PM, Florian Binder wrote:
>
> In the gc log it seems that there is always running the same garbage
> collection.
> For example:
> $ zcat out_err.log.gz | grep 20480M | tail
> 2011-11-28T19:13:19.482+0100: [GC cleanup 14515M->14515M(20480M),
> 0.1370060 secs]
>    [ 16394M->12914M(20480M)]
>    [ 16394M->12914M(20480M)]
> 2011-11-28T19:17:12.509+0100: [GC cleanup 15582M->15582M(20480M),
> 0.1387230 secs]
>    [ 16394M->12914M(20480M)]
>    [ 16394M->12914M(20480M)]
>    [ 16394M->12914M(20480M)]
> 2011-11-28T19:21:06.089+0100: [GC cleanup 12978M->12978M(20480M),
> 0.1344170 secs]
>    [ 16394M->12914M(20480M)]
>    [ 16394M->12914M(20480M)]
>
> Therefore I assume this might be a bug ;-)
> You can download the whole log at: http://java4.info/g1/out_err.log.gz
>
> I don't think that we still have the logs of that very long gc, but I will
> have a look for it at monday.
> Furthermore I do not think that we have logged there much details of the
> gc. But I know that this happend on a very special server, which contains
> more the 30gb of references (yes, just references to other objects). If we
> run it with CompressedOops we reduce the memory useage to nearly 50%.
>
> Flo
>
>
>
> Am 10.12.2011 01:34, schrieb Srinivas Ramakrishna:
>
>
>
> A couple of things caught my eye....
>
>   On 12/05/2011 02:19 PM, Florian Binder wrote:
>> >> I assume you know when the data will be dropped and reinitialized,
>> >> right? Can you do a Full GC (with System.gc()) after you re-initialize
>> >> the heap. This typically helps a lot with CMS.
>> > Yes, this is exactly what we are trying at this time ;-)
>> > Either this Full-GC pauses are very long (60-300 seconds) they are much
>> > shorter than after a CMS-failure (when they take sometimes more than
>> > 4000 seconds ;-)) and we can define when they occure.
>>
>
> A full GC that takes an hour is definitely a bug. Have you logged that bug?
> Or at least share the GC log? What's the version of the JDK that this
> behaviour
> was seen with?
>
>  ...
>
>>  > Furthermore this should not increase the total-heap-space:
>> > http://java4.info/g1/all.png
>>
>
>  Remember that jconsole asynchronously samples the heap, whose size is
> read "with possible glitches". Rather, you should probably rely on the GC
> log in order to assess
> the heap size after each GC event, rather than the asynchronous samples
> from
> jconsole. I myself have not had the chance to look at yr GC logs to see
> what that indicated wrt the size of Eden and of the Heap.
>
> -- ramki
>
>
>
>
> _______________________________________________
> hotspot-gc-use mailing listhotspot-gc-use at openjdk.java.nethttp://mail.openjdk.java.net/mailman/listinfo/hotspot-gc-use
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.java.net/pipermail/hotspot-gc-dev/attachments/20111212/79b2b01a/attachment.htm>


More information about the hotspot-gc-dev mailing list