Low-Overhead Heap Profiling

[Thomas Schatzl]

Hi,

On Wed, 2017-06-28 at 09:51 +0200, Robbin Ehn wrote:
> Hi,
> 
> On 06/21/2017 10:45 PM, JC Beyler wrote:
> > 
> > Hi all,
> > 
> > First off: Thanks again to Robbin and Thomas for their reviews :)
> > 
> > Next, I've uploaded a new webrev:
> > http://cr.openjdk.java.net/~rasbold/8171119/webrev.06/
> > 
> > Here is an update:
> > 
> > - @Robbin, I forgot to say that yes I need to look at implementing
> > this for the other architectures and testing it before it is all
> > ready to go. Is it common to have it  working on all possible
> > combinations or is there a subset that I should be doing first and
> > we can do the others later?
>
> I will change my mind here and proposed a different approach.
> Right now the only compiler platform change you have is for the
> FastTLABRefill.
> FastTLABRefill works for all gc except G1 (CMS is deprecated) but
> only in C1 compiled code.
> The performance angle of this is not really relevant, since then you
> will have C2 compiled code.

Afaik C2 compiled code also uses the stubs for TLAB refill/inline
allocation too (I may remember wrongly).

> So I suggested that we should remove FastTLABRefill completely, I
> will try to makes this happen but most likely not before my long
> vacation.
>
> That means you can avoid all platform specific code (as patch is
> now), correct?
> 
> If so just keep the x86 support in there and we remove that if I
> succeed, plus you can support all platforms when
> FastTLABRefill=false.
> 
> Sounds good?

Not sure about it is easy to remove FastTLABRefill, but that may be
something to discuss in more detail. Just noting that while the runtime
call can be negligible on large heaps (with large TLABs) it might not
be on small ones (or you have many Java threads which causes your TLABs
to get smallish due to how TLAB size is calculated iirc).

E.g. Serial GC may still be used a lot on embedded platforms which tend
to use smaller heaps and benefit the most from this optimization.

Now one can always automatically disable FastTLABRefill if you want
this kind of heap monitoring (and it does not work with
FastTLABRefill).

Thanks,
  Thomas