Beowulf And Digital Signal Processing - Success!

[Gerry Creager]

Kragen wrote:
> 
> On Tue, 29 Sep 1998, Huntress, Gary B. wrote:
> > With 10 nodes (9 worker nodes) I observed a speedup of 8.5 with an 8k
> > FFT.  I'm very pleased with these results, and I just thought I'd share
> > a success story.  I strongly suspect that this will lead to more use of
> > clusters in the DSP area (anyone doing this?  I'd love to hear from
> > you).
> 
> I would think that, if you had a lot of DSP work to do, it'd be cheaper
> to do it with DSPs than with 486s.  Something like the TigerSHARC
> card.  Doing matrix multiplies, vector math, or Fourier transforms is
> likely to be a heck of a lot faster with DSPs than with x86s -- like a
> factor of 10 for CPUs that cost the same amount.

Sometimes, you use what you have.  If the budget doesn't allocate
dedicated DSPs, you use discarded 486's... I had to build from discarded
486's here, as well, to develop a proof of concept system: The local
nay-sayers didn't believe it'd work and had to be shown.  Point of fact:
The cluster benchmarks slower than my dedicated linux research system
(PII-266) but demonstrates a performance increment when compared to one
or two 486's, allowing even the Luddites to imagine an increment with
"real" CPUs.

Gerry Creager
Mapping Sciences Laboratory
Texas A&M University