RFR (M): 8047328: Improve memory usage for cards in SparsePRTEntry

[Erik Helin]

On 2016-05-10, Thomas Schatzl wrote:
> Hi all,
> 
> On Mon, 2016-05-02 at 16:04 +0200, Thomas Schatzl wrote:
> > Hi all,
> > 
> >   can I have reviews for this change that cuts SparsePRTEntry usage
> > by
> > half (asymptotically at least :) )?
> > 
> > This change is based on a change from A. Sjöberg (which will be
> > credited by me), where instead of having special -1 entries in the
> > SparsePRTEntry array of cards, use a dedicated next-pointer for
> > storing
> > where to put the next card.
> > 
> > This allows us to internally use uint16_t's instead of CardIdx_t's
> > (which are 32 bit) for storing.
> > 
> > This also limits the max region size to 32M: there is an assert that
> > checks that - unfortunately it is not possible to have a
> > STATIC_ASSERT
> > here, as HeapRegionBounds::MAX_REGION_SIZE is private.
> > If you think that it is worth to have a static assert, at the cost of
> > exposing that member (or has another idea on how to do that without
> > adding too many dependencies somewhere), I would be happy to change
> > the
> > code accordingly.
> > 
> > This reduces memory usage of sparse prt entries, and also somewhat
> > speeds up searching and iterating entries.
> > 
> > There are no significant throughput gains for that change though, but
> > the memory savings (particularly on large heaps) are nice to have.
> > 
> > (Even in the worst case, i.e. very small default number of cards in a
> > SparsePRTEntry we are not worse than before in that regard).
> > 
> > CR:
> > https://bugs.openjdk.java.net/browse/JDK-8047328
> > Webrev:
> > http://cr.openjdk.java.net/~tschatzl/8047328/webrev/
> > Testing:
> > jprt, vm.gc, perf benchmarks
> 
>   Erik had a look at the change offline and suggested to remove some
> changes that are not absolutely necessary and detract from the
> important points.
> 
> New webrevs at
> http://cr.openjdk.java.net/~tschatzl/8047328/webrev.0_to_1 (diff)
> http://cr.openjdk.java.net/~tschatzl/8047328/webrev.1 (full)

Looks good, Reviewed.

Thanks,
Erik

> Testing: another jprt run.
> 
> Thanks,
>   Thomas
>