RFR (S): 8003237: G1: Reduce unnecessary (and failing) allocation attempts when handling an evacuation failure

[Thomas Schatzl]

Hi all,

  just fyi, I had to update the webrev because of changes in the 8073052
patch.
Since there were no reviews of these changes yet, I updated in-place.

There are no functional changes.

Also, if somebody already had a look at the changes, and the question
why G1ParScanThreadState::allocate_in_next_plab() only uses failed PLAB
allocation as indicator that heap is full came up, the later patch
JDK-8067336 "Allow that PLAB allocations at the end of regions are
flexible" will allow that PLABs can get quite small as they are sized
flexibly.

I wanted to avoid introducing some even more arbitrary heuristic to
determine when the heap is full now which will only a bit later be
superseded again anyway.

In any case current G1ParScanThreadState::allocate_in_next_plab() should
try harder than the previous one by always trying an inline allocation
even if it cannot allocate the object (supposedly smaller than a PLAB)
using a new PLAB.

Thanks,
  Thomas

On Mon, 2015-08-03 at 14:43 +0200, Thomas Schatzl wrote:
> Hi,
> 
>   this change is based on the latest version of the change for 8073052
> that is out for review, and is still longing for another reviewer :)
> 
> (I am okay with pushing it with only a single "R"eviewer since it is
> only refactoring, but if somebody can find the time...)
> 
> Thanks,
>   Thomas
> 
> On Mon, 2015-08-03 at 14:35 +0200, Thomas Schatzl wrote:
> > Hi all,
> > 
> >   can I have reviews for this small change that improves evacuation
> > failure handling by short-cutting copy_to_survivor_space() when there is
> > no more space left?
> > 
> > In effect this makes the parallel phase during GC have the same
> > performance as if there is no evacuation failure, and I could not
> > measure regressions in the normal case either.
> > 
> > The change implements two things: first, notice when there is no more
> > space left in the last gen (i.e. old), and if so, short-cut
> > ParScanThreadState::copy_to_survivor_space(). The other change is to not
> > try to ask the free list for new regions if it's empty to avoid taking
> > the lock in the time window between the first few evac failures and when
> > space is completely empty.
> > 
> > There is still the somewhat heavyweight cleanup phases after evacuation
> > failure.
> > 
> > CR:
> > https://bugs.openjdk.java.net/browse/JDK-8003237
> > Webrev:
> > http://cr.openjdk.java.net/~tschatzl/8003237/webrev/
> > Testing:
> > jprt, lots of performance measurements with and without this change
> > compiled in
> > 
> > Thanks,
> >   Thomas
> > 
> > 
> 
>