Re: How do you setup an Alpha cluster

[Greg Lindahl <lindahl cs virginia edu>]

> A bunch of us at the STA Branch of the NASA Ames Research Center are
> sold on the idea of using an array of Alphas running on PCI bus
> motherboards to do scientific computations.

Doing this is about the same as setting up an x86 PC-based cluster.

> For example: would fast Intel or AMD chips running Linux be a better
> approach.

Benchmark your code to find out. SPEC shows a huge advantage for the
Alpha, but while you can buy the PGI compiler for Linux/x86 and get
that actual speed, on the Alphas, for some codes, you might run only
50% as fast as you would expect from SPEC. On the other hand, your
code might be limited by main memory bandwidth, which is a whole
'nother ball of wax.

> 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?

None. Well, non-SMP Linux is more stable than SMP Linux, but everyone
should realize that.

>  Is using an Ethernet
> message passing approach the best basis for an Alpha cluster running parallel
> code?

Probably. A benchmark will tell you if your applications run OK with
fast ethernet. And fast switched ethernet is much cheaper than the
next level of performance (myrinet or gigabit ethernet.)

>  What's the optimal top CPU speed for an Alpha cluster? 

You should find prices for a variety of speeds and plot speed verses
the price of one complete system. You'll discover that it's a curve,
and you want to buy at the "knee". When I bought, 533 mhz was much
less expensive than 600 mhz; I could get much more total power by
buying 533 mhz machines.

You do the same thing when pricing an x86-based cluster.

-- greg