RFR (M): 8153503: Move remset scan iteration claim to remset local data structure

[Thomas Schatzl]

Hi,

On Wed, 2016-04-13 at 13:27 -0400, Derek White wrote:
> Hi Thomas,
> 
> Small comment at the end...

[...]

> > > I'll look at this more on Monday.
> > >    - Derek
> 
> Sorry for dripping this out. I spent a lot of time trying to 
> understand what G1ScanRSClosure::doHeapRegion() was really doing - 
> what it was iterating over, how it was filtering out the claimed 
> cards, etc.  If that iteration were done anywhere else I'd suggest 
> separating it out into a "claimed" iterator class (but it's not done
> anywhere else).

Not sure what you mean, but probably something like SubTasksDone or so.

The problem is that in this case, you want other threads to help each
other when one region takes too long. So really every thread wants to
iterate over all heap regions.

Further, there is this matter that parts of the work should only ever
done by a single thread if possible (adding the region to the dirty
cards regions) or must be (code root scan).

One can of course separate the tasks (update rs, scan rs and scan code
roots), but I do not really want to address this here. There is the
separate issue JDK-8153505 already.

> One suggestion to existing code:
> 
> g1RemSet.cpp, Line 235 (etc):
> 
> Rename "jump_to_card" => "claimed_cards". The "jump" term pre-dates
> the 
> current parallelization scheme and now isn't very meaningful.

I think nothing changed conceptually in this part. I will see if I can
find a better name for that variable though. :)

There are much better ways to handle work distribution for scan rs, but
in this CR (and probably for JDK 9 FC) I would prefer if we did not
change how claiming work works.

Thanks,
  Thomas