[Beowulf] Cell programming

[Joe Landman]

Hi Tim

Tim Wilcox wrote:
> I have been coding on the Cell for just over a month now.  Nothing serious, 
> just getting subroutines to run on a single SPU.  This turns out to be very 
> easy.  Now I am looking at parallel programming between the SPUs and this 
> seems much more difficult.  The API, as far as I have read it, does not have 
> nice routines for message passing between the SPUs, you have to set up your 
> own memory transfers or address remote memory directly using the MFC.
> 
> Running independent codes on each SPU is pretty easy but only addresses the 
> class of ideally parallel problems.  I would like to encourage a discussion 
> on techniques of how to split and parallelize a problem with 
> this "dual-layer" parallel architecture(MPI and MFC).  
> 
> It seems to me a good starting point is to divide a problem, say a CFD, into 
> larger sections at the MPI layer and then a smaller division of the subset on 
> the individual Cell processor.  This poses the issue of message passing 
> between disparate SPUs.  Any input on the characteristics of how this may 
> work and perform?

This sounds a great deal like the multi-layer parallel approach the 
MPI-HMMer team have used.  Coarse grain layer subdivision with MPI, fine 
grain layer uses SSE2.  Net effect is nice speedup in parallel, and good 
single CPU performance.

I have some nice data on this that I had run over the last few months 
using MPI-HMMer.  Seeing ~40x on 64 CPU cores over single CPU core 
(depends on HMM and database size, as this significantly impacts the 
communication to computation ratio, as well as the work the scheduler 
has to do).

> 
> Tim Wilcox
> Terra Soft Solutions
> 505-239-4600
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf


-- 

Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: landman at scalableinformatics.com
web  : http://www.scalableinformatics.com
phone: +1 734 786 8423
fax  : +1 734 786 8452 or +1 866 888 3112
cell : +1 734 612 4615