Request for Cluster Recommendations

[Michael McDonald]

Thanks for the reponses.  I should have
made it clearer what my applications will
be but you all doped it out pretty well.  I'm
setting it up for applications in computational
chemistry and physics.  A lot of eigenvalue
and fourier computations.  The codes
use the netlib LAPACK and blas libraries
and FFTW comes up a lot.

For the most part, my codes have been data
parallel and have invlolved broadcasting
parameters and merging results at the end of
distributed serial computations.  They have
involved many evaluations of small matrices
and data sets and I get better speedups with
simultaneous serial executions.  Much of the
work I'm looking towards would make use
of a grid approach in the style of seti at home
or the factoring projects.

I would expect bursts of communication
separated by periods of computation;
Overall, communication wouldn't be 
so much of a bottleneck, but I'd like it
to be fast when it does occur.  If Andy
Sporner's 30% figure holds up, I think
local disks as buffers would allow the
network access to be smeared out.  I've got
40% in mind as an upper performance limit
for ethernet due to collisions, but I can't back
that up.  Local buffers seem to allow for
scheduling comm. so as to avoid collisions.

Starting out with 100Mbit may make sense as
a cheaper/simpler startup - no NIC
purchases.  After running some simulation
and benchmarking with the apps, upgrading to Gigabit wouldn't be an undue burden.
Some initial attempts at estimating
communication & compute demands would
be in order. Aside from the fiber options, is the cabling the same for 100Mbit and Gbit?

Thanks to all,
Mike McDonald