>>>>> "Richard" == Richard Hayes <[EMAIL PROTECTED]> writes:

Richard> Dear list, I am think of building a prototype supercomputer
Richard> and if it works I will build a much larger system.

To do what?  Do you need shared memory, or are you thinking Beowulf
cluster?

The key things are the scalability of the algorithms you want to run.
This leads to trade off between a few high-power processors, or lots
of slower processors.  Also, how important is floating point
performance?

If you go for a cluster, then memory bandwidth and interconnect speed
are the two big issues.  If you're using MPI then drivers are
available for infiniband and myrinet interconnects that run in
user-space, and avoid the latency overhead of going through the Linux
kernel.  In general, the newer IA64 machines have much better memory
bandwidth than commodity hardware.  But it all depends on your
application.


Richard> What is the best way to get cluster for testing.  I know that
Richard> Stone Soupercomputer cost zero but really I have two
Richard> questions.

Richard> 1. How does the AMD64 stack up in cost / performance compared
Richard> to the say 1.8 Duron?

Don't know -- I use IA64...

Richard> 2. How do you create SAN / NAS type storage for large data
Richard> sets?

You can buy then from SGI etc.
-- 
SLUG - Sydney Linux User's Group Mailing List - http://slug.org.au/
Subscription info and FAQs: http://slug.org.au/faq/mailinglists.html

Reply via email to