[Beowulf] choosing a high-speed interconnect

[Chris Sideroff]

Thanks for all the excellent replies.  The consensus seems to be carry
out some performance profiling with the current hardware and compare
with a high-speed network cluster.  The former which I am in the process
of doing and the latter which I will try to do.

More specifically, I believe the most important testing for our cluster
will be Fluent's scalability and sensitivity to the network.  The reason
I say this is because there are multiple users (~6-8) running large
Fluent jobs (1-10 million cells) with various solvers, which have
different CPU and memory requirements.  While the in-house code is run
by one person, a lot less frequently and (for other reasons) currently
does not use more than 8 processors.

Following rgb's suggestions, I will be difficult to analyze Fluent at
the code level since we don't have the code but it has some built in
monitors that I can use combined with some Linux tools.  When the time
comes to scale the in-house code to >8 procs we will have greater
flexibility tuning it. BTW, thanks for the parallel programming
reference.

I did find some benchmarks on Fluent's website which indicate that
Myrinet definitely scales better than GigE but I still want to carry out
my own benchmarking.  If anyone has any experiences benchmarking their
clusters using Fluent feel free to supply your thoughts.