[m-rev.] diff: upgrade to Boehm GC 7.2alpha4

[Paul Bone]

On Thu, Mar 11, 2010 at 04:08:24PM +1100, Paul Bone wrote:
> On Wed, Feb 24, 2010 at 05:44:55PM +1100, Julien Fischer wrote:
> >
> > I have omitted the full diff of the boehm_gc directory - it's gigantic,
> > and not very interesting.  (It's in /home/mercury/juliensf/tmp/BOEHM_GC.UPGRADE
> > for those that are feeling brave.)
> >
> > I suspect the following will also fix some build problems on OpenSolaris
> > and some of the BSDs, but I haven't checked yet.
> >
> > -------
> 
> Thanks Julien,
> 
> This has improved preformance in general.  (The ICFP2000 raytracer takes about
> 60 seconds now compared to 83 seconds previously for the same render).
> Although more formal testing is required.

> This has also improved scalability wrt Parallelism.  With parallel marking
> disabled the and when using 8 threads on a Intel i7 (4 cores X 2 threads).  The
> raytracer runs in 35 seconds (it runs in 60 seconds for the asm_fast.gc grade).
> Previously with parallel marking disabled it only minimally cut the execution
> time down, there's a graph on my desk.
> 
> With parallel marking enabled this drops down again to 19 seconds.  3 times the
> speedup for 4 cores isn't bad for Boehm + Mercury.  Of course I think there's
> plenty of room to improve.
> 
> Disclaimer: I ran these tests to satisfy my curiosity, they arn't exactly
> scentific but the results are plausable and are probably representative of the
> true results.
> 

Actually these result can't be trusted.  I don't remember which machine I did
the 'before' test on.  When I sent this I thought I did it on the i7.  I'll do
more testing.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 489 bytes
Desc: Digital signature
URL: <http://lists.mercurylang.org/archives/reviews/attachments/20100311/e359fda1/attachment.sig>