[m-dev.] Boehm parallel marking performance.

[Paul Bone]

On Tue, Jun 02, 2009 at 01:43:11PM +1000, Peter Wang wrote:
> 2009/6/2 Paul Bone <pbone at csse.unimelb.edu.au>:
> >
> > I've tested the performance of the compiler (using speedtest) with and
> > without parallel marking enabled in the garbage collector.  Here are the
> > results.
> >
> > Each test was performed with 50 samples, each sample attempts to compile the
> > dozen current largest (in terms of bytes) modules in the mercury compiler.
> > This was run on taura.  I've connected all the statistics generated by
> > speedtest, only the 'wall time' measurement is shown below.
> >
> > Parallel mark improves the speed of the compiler by roughly 6.5 percent when
> > running in the asm_fast.gc.par grade.  This option is enabled by default in the
> > parallel version of the garbage collector.
> 
> I assume you deleted @ENABLE_BOEHM_PARALLEL_MARK@ from Mmake.common.in.

Instead I commented out the line that set it for linux in configure.in

> It would be interesting to see the improvement from parallel marking in
> non-.par grades, and if enabling parallel marking at compile time but not
> at runtime has any adverse effects. (Set GC_NPROCS=1)

I thought about this breifly, intuitivly I think the synchronization required
might make it slower, since without parallel mark in a sequential program
synchronization is not needed at all.

I don't understand why you would want to test parallel marking at compile time
but then disable it at run time.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 197 bytes
Desc: Digital signature
URL: <http://lists.mercurylang.org/archives/developers/attachments/20090602/6f3e7f8f/attachment.sig>