Region-level compaction and related considerations for G1 to better facilitate H-objects

[Thomas Schatzl]

Hi Tao,

On Fri, 2015-06-26 at 17:46 -0700, Tao Mao wrote:
> Hi there,
>
> Recently in using G1 on our realtime-focused production servers, I
> noticed some full GCs were triggered by humongous allocation failure
> due to fragmentation of free regions while it's still having enough
> space in number.
>
> I'm not suggesting that we move H-objects around. Just wonder if we
> have the following or similar in place or plan to have. e.g.,

In such cases, moving around (small) humongous objects might actually be
a good idea (compared to a full gc).
> 
> (1) For evacuation, we grab free regions closer to the start of heap
> and/or try to grab free regions with fewer neighboring free regions.

1a) has been implemented in some 8u release, either u20 or u40. If I
remember correctly, new old gen regions (that includes humongous) are
allocated from the end of the heap downwards, new young regions from the
start of the heap.

The second option has not been explored (try to grab free regions
according to other policies), but seems straightforward (but some work)
to do. Check out G1CollectedHeap::new_region().

Controlling the destination regions only solves half of the problem
though, but a useful mechanism also requires control of the source
regions: in a given evacuation there are often a lot of almost equally
expensive regions to collect, so any change according to e.g. "try to
create as much as possible contiguous free areas" might help a lot
without a lot of performance impact.

> (2) Humongous object allocation tries to find the most suitable
> location, i.e., least waste of its neighboring free regions after
> H-allocation.

Check out G1CollectedHeap::humongous_obj_allocate(). The entry point
that might be most interesting would be to change
HeapRegionManager::find_contiguous_empty_or_available().

> Is there something along this line to fight region-level
> fragmentation, or related fixed bugs? If so, any pointers?

Above mentioned is what I know in that area.

Thanks,
  Thomas