Very long duration of Finalize marking in the Remark stage of G1GC

[Thomas Schatzl]

Hi,

On Wed, 2015-10-14 at 14:12 +0300, Ivan Shramko wrote:
> Hi Thomas!
>
> Did you mean humongous objects? We don't have them. Grepping gc log
> with "humongous" doesn't show anything.

Yes, I meant humongous objects. In the log snippet you gave, it did not
indicate any amount of these objects too, e.g.

>           [Humongous Register: 0.0 ms]
>           [Humongous Reclaim: 0.0 ms]

which typically indicates none. However both actions are typically very
fast, so 0.0ms is not really unusual even in presence of some. In your
log snippet I saw that some Humongous Register phases are 0.1ms, but
that may just be some random variation.

The options I gave you will make that clear (although I saw that you
already have PrintAdaptiveSizePolicy enabled) because they will print
the number of humongous objects directly.

If you already mention that you are sure you do not have them, the
mentioned issue cannot be the cause anyway.

> I attached part of gc log with young gcs before long finalize marking.
> 
> I have added JVM options you asked and will post log after next long
> finalize pause.

Maybe the new log will yield some new information. Not sure.

Thanks,
  Thomas