RFR (M): 8034842: Parallelize the Free CSet phase in G1

[Thomas Schatzl]

Hi all,

  due to recent pushes the change did not apply cleanly any more.

http://cr.openjdk.java.net/~tschatzl/8034842/webrev.0_to_1/ (diff)
http://cr.openjdk.java.net/~tschatzl/8034842/webrev.1/ (full)

contain changesets that do.

Thanks,
  Thomas

On Mon, 2016-05-02 at 16:04 +0200, Thomas Schatzl wrote:
> Hi all,
> 
>   can I have reviews for the free cset phase for G1?
> 
> We observed that the part of the free cset phase that takes the most
> time is counting the number of occupied cards.
> 
> While earlier changes focused on decreasing the time to do that, this
> change parallelizes this work, which is unfortunately a bit
> complicated, but gives the expected results.
> 
> The change splits out the parallelizable part from the serial (pre-,
> mid- and post-) part:
> 
> During the first serial part the change creates work lists for the
> parallel part
> (G1FreeCollectionSetTask::setup_for_parallel_processing()), which is
> in
> itself split into operations that need to be applied to young and old
> gen remsets (doing the slow freeing of the remsets essentially;
> G1FreeCollectionSetTask::do_parallel_work()), and some additional
> serial "mid-" part (G1FreeCollectionSetTask::do_serial_work()) that
> is
> performed in parallel to the rem set freeing.
> 
> After that, the change updates the global data structures in another
> post-serial phase (G1FreeCollectionSetTask::complete_work()).
> 
> Probably it's easiest to review this by comparing the main loop of
> the
> old G1CollectedHeap::free_collection_set() method beginning in line
> 4736 of g1CollectedHeap.cpp with the code
> in G1FreeCollectionSetTask::do_serial_work().
> 
> CR:
> https://bugs.openjdk.java.net/browse/JDK-8034842
> Webrev:
> http://cr.openjdk.java.net/~tschatzl/8034842/webrev/
> Testing:
> vm.gc, many many benchmarks
> 
> (This is a rather old patch, having been used locally and many some
> internal benchmark runs quite a bit)
> 
> Thanks,
>   Thomas
>