State of parallel GC?

[Simon Marlow]

Bryan O'Sullivan wrote:
> Ravi Nanavati wrote:
> 
>> It isn't an immediately pressing need a Bluespec, but we do get asked
>> every once in a while about parallelizing the Bluespec compiler. 
> 
> Roshan James was working on parallel GC at MSR Cambridge last year. 
> Unfortunately, it's not an internship-sized project.
> 
>> One
>> of the technical reasons we're holding off on that front is that it
>> wouldn't win as much as people expect (because garbage collection is
>> usually a decent fraction of a typical run), so we'd be interested if
>> the state-of-play in that area changed.
> 
> I think that the need for it will become more pressing among GHC's 
> developers as the NPH work matures, as the current GC reduces 
> performance even on two CPUs.  The need to publish good numbers is 
> always a helpful motivator :-)

I tried it again recently, and for some reason I have yet to fathom, my 
parallel GC doesn't actually run any faster than single-threaded GC.  That 
is, it divides the work equally between the processors, all processors are 
kept busy, but the total elapsed time is the same or greater than running 
the GC single-threaded.  It seems unlikely that I've reached the limit of 
the memory bandwidth, but that's what it looks like.

Rest assured that this is a high priority for us too.  I have PAPI running 
on my laptop, and I'll keep investigating, hopefully the problem will 
reveal itself at some point.

Cheers,
	Simon