new SGI Origin & Onyx 3k?

Douglas Eadline deadline at
Thu Jul 27 05:08:55 PDT 2000

On Wed, 26 Jul 2000, W Bauske wrote:

> Matthias Troyer wrote:
> > 
> > W Bauske wrote:
> > > Greg Lindahl wrote:
> > > > > Exactly how often do you think people want to run 1000 nodes on
> > > > > a single problem? When was the last time you did that?
> > > >
> > > > For the customers I'd like to have, frequently.
> > > >
> > >
> > > Note the "Like to have". Also, you don't address the first
> > > question so I'll assume you agree people don't do that unless
> > > it's for benchmarks from your statement.
> > 
> > We are actually routinely running production codes on several hundred
> > nodes, and are devloping codes that will need 1000 nodes AND a fast
> > network for a single problem. Our 500 CPU Beowulf cluster is just not
> > the right choice for some problems.
> > 
> Do you consider your system environment "normal"?
> Some people do use very large systems. They are not in the 
> majority however. Just not that many companies are willing
> to spend the $$$$$$$$ to keep such systems around.
> I also ask you what sort of cpu's you use. If you're even
> one generation back, you should just upgrade the number
> of cpu's you have to the latest generation. It makes me chuckle
> when someone tells me they have 1000+ processors and they turn
> out to be 4-5 years old.

Actually, if you are writing the checks, it all depends on
price-to-performance and accounting issues. In addition, faster
processors are not always best (sounds crazy - read this paper

> Also, which vendor you choose can make a significant difference
> on where 500 processors perform. 500 Intel PIII's are not the
> same as 500 21264 Alphas, unless your work involves much integer
> work. 

For many customers, it is how fast your code runs and how much
the box costs (and how much it costs to keep it running). To those
that spend money it is price-to-performance and total cost of ownership
(TCO). This is largely due to the fact that purchasers have limited
funds and must maximize performance within this limit.

If absolute performance is the issue then funds are normally
not as limited and the latest and greatest is used.

In the real world, sometimes Alphas are better and sometimes PIII's
are better and sometimes a bunch of hamsters running in a wheel are
better, and sometimes Beowulf clusters are not the answer. It all depends
on the application.


Paralogic, Inc.           |     PEAK     |      Voice:+610.814.2800
130 Webster Street        |   PARALLEL   |        Fax:+610.814.5844
Bethlehem, PA 18015 USA   |  PERFORMANCE |

More information about the Beowulf mailing list