[geeks] Massively parallel... cheap?

[Sridhar the POWERful]

On Sat, 23 Mar 2002, David Passmore wrote:

> I am in need of some recommendations. I have a project in mind that will
> require a machine (or cluster) with some pretty beefy capabilities:
>
> * 64+ processors (the more, the better!)
> * Several gigabytes of memory -- this is basically a big disk cache
> * ability to address 1TB+ of disk
>
> The computation this system will be performing is most likely going to be
> simple integer math and comparisons-- *maybe* floating point but I doubt it.
> The working data set will be *huge*, several gigabytes in size, with
> somewhere between 10000 and 100000+ data points being computed at any given
> 'time'. Think about a talk similar to simulating the movement of a million
> stars and that's pretty much what I'm up against, which is why it has to
> execute as many operations in parallel as possible. If someone has something
> brilliant to say here, I'd sure love to hear it.
>
> The real catch: this has to run in my basement, or in colo, so the Cray on
> eBay is out of the question. ;) What can folks suggest in terms of a system
> or cluster to run this kind of task on?

Is this going to be processor-bound or I/O-bound?  If it's I/O bound, get
a used S/390.  They are capable of moving *humongous* amounts of data
without breaking a sweat.  If it's processor-bound, get a bunch of small
alphas and set up a VMScluster.  Just make sure you do it with a really
good robust network.

Peace...  Sridhar