Basic question

[Craig Tierney]

> I have a maybe silly question to Beowulf experts.
> We are planning to set up a small cluster of ~10 nodes during the next
> year, and if the
> results (in terms of speed) are ok, we think to enlarge it. Currently we
> are using CRAYs
> and MPI.
> 
> Now to my questions
>     1) which linux-distribution is best for beowulf (redhat, suse, ...)

Which ever you like best.  We happen to use Red Hat, but we have experience
with it.  You can get into big holy war over this, but use what you like.

> ?
>     2) what to you think is better PCs with one or more processors ?

Dual CPUs are nice because you don't have to worry about the infrastructure
of a second node.   Communication is better between the two nodes (or
at least it should be) than two nodes connected with "your favorite
interconnect".

Downsides:  You may not be able to install as much memory per node
as you would like.  Only a few x86  boards can do > 1 GB RAM.  Also, if you 
want Athlons, there are no dual CPU chipsets (yet).  You also do not get
100% of the second CPU.  Memory bandwidth will usually get in your way.

In my testing dual CPU packaging is better than single CPU for 
price vs. performance.  Your mileage may vary.

>     3) how's about connections between this PC's (Gbit/s,....) ?

What kinds of problems are you solving?  Is latency and/or bandwidth
important?  For codes here (weather and atmosphere finite difference codes)
Compaq Alpha XP1000s + Myrinet provide very good scaling and performance.  On
these types of codes I was seen a 3x performance over the codes on
a T3E.  Again, your mileage may vary.

>     4) any additional hints ?

Go real all the archives.  There is alot of good information there about
these kinds of questions and hardware that people are using in their
clusters.

You didn't ask about software.  That is another can of worms.

Good luck.

Craig

-- 
Craig Tierney (ctierney at hpti.com)
phone: 303-497-3112