G1 root cause and tuning

[Thomas Schatzl]

Hi all,

On Mon, 2015-03-30 at 20:41 -0500, charlie hunt wrote:
> Hi Jenny,
>
> One possibility is that there is not enough available contiguous
> regions to satisfy a 300+ MB humongous allocation.
>
> If we assume a 22 GB Java heap, (a little larger than the 22480M shown
> in the log), with 2048 G1 regions (default as you know), the region
> size would be about 11 MB. That implies there needs to be about 30
> contiguous G1 regions available to satisfy the humongous allocation
> request.
>
> An unrelated question … do other GCs have a similar pattern of a
> rather large percentage of time in Ref Proc relative to the overall
> pause time, i.e. 24.7 ms / 120 ms ~ 20% of the pause time.  If that’s
> the case, then if -XX:+ParallelRefProcEnabled is not already set,
> there may be some low hanging tuning fruit. But, it is not going to
> address the frequent humongous allocation problem.  It is also
> interesting in that the pause time goal is 2500 ms, yet the actual
> pause time is 120 ms, and eden is being sized at less than 1 GB out of
> a 22 GB Java heap.  Are the frequent humongous allocations messing
> with the heap sizing heuristics?

While I have no solution for the problem we are aware of these problems:

- https://bugs.openjdk.java.net/browse/JDK-7068229 for dynamically
enabling MT reference processing

- https://bugs.openjdk.java.net/browse/JDK-8038487 to use mixed GC
instead of Full GC to clear out space for failing humoungous object
allocations.

I am not sure about what jdk release "JRE 1.17 update 17" actually is.
From the given strings in the PrintGCDetails output, it seems to be
something quite old, I would guess jdk6?

In that case, if possible I would recommend trying a newer version that
improves humongous object handling significantly (e.g. 8u40 is latest
official).

Another option that works in all versions I am aware of is increasing
heap region size with -XX:G1HeapRegionSize=<X>M, where X is 8/16 or 32;
it seems that 4M region size has been chosen by ergonomics.
Start with the smaller of the suggested values.

Thanks,
  Thomas