Re: How do you setup an Alpha cluster

["Street Dreams" <strdream iserv net>]

> The basic idea is our engineer/scientists would use their desktop
machines for code development or pure scalar computations and farm out
their extremely parallel algorithm CPU intensive work to the Alpha cluster.
 The primary scientific application would be trajectory and computational
fluid dynamic simulations of spacecraft entering planetary atmospheres.
> Great minds work alike!  I understand that a bunch of people at Los
Alamos and other places have built up Alpha clusters based on PCI bus
motherboards and the special effects for Titanic came from the same sort of
architecture.  

That system isn't really a parallel cluster, in the Beowolf sense.  A set
of frames is assigned to each of some 200 systems in a batch render job.

if it weren't for DiCaprio I just might see the movie (or rent and fast
forward to see the FX :) )

>I'm a newbie humbly seeking advise from people who have actual experience
building these systems.  Is this basic concept a dumb idea?   For example:
would fast Intel or AMD chips running Linux be a better approach.  Is using
Linux a dumb idea?  Should I (shudder) be using only Windows-NT [Jeff: HA
HA HA]?  What are the hidden pitfalls one encounters with an Alpha cluster
that do not appear until after the system has been built and running for a
few weeks?  Is using an Ethernet message passing approach the best basis
for an Alpha cluster running parallel code?   What's the optimal top CPU
speed for an Alpha cluster?  For example if the Ethernet latency is the
primary data bottleneck then a cluster based upon the more expensive 800
Mhz. CPUs might perform no better than 533 Mhz. CPUs.  Our goal is to
maximize CPU-cycles/dollar.  We understand that there are faster
architectures out there >(Cray T3D) but we can not afford them. 

The Titanic FX group, Digital Domain, determined that using Intel
compatible machines would require twice as many systems to do the same job
for them, and SGI had as good performance as Alpha, but too pricey.  And
rendering is FPU intensive, like many scientific computations, so you
should get similar results.  Integer performance is sharp too, esp. at the
64 bit width data.

As for bottlenecks, keep in mind the relative number of message passing
required for your average job.  The messages have to  cross the ethernet to
other systems, so maximizing the number of ethernet busses is good, and
keep the minimum network minimum speed to 100Mbps.

If you can get 533's they're the best for money but I hear they're not made
any more, so make sure you can upgrade to faster chips with the boards
you're getting.  Prices should drop soon.

The Beowolf cluster is most bang for buck intensive that I know about.  I
understand that RedHat's Extreme Linux package simplifies the software
setup.  I know a guy that's building a cheap 16 node 486-castaways cluster
as a school project, but I haven't done this.

Get a hold of Linux Journal's Clustering issue.  It covers the basics that
I don't remember at the moment.  Go to www.ssc.com to order or contact
them.

Jeff D.