[Fwd: Benchmarking Beowulf]

[R. Munk Larsen]

Greg,

> 
> No; there are real problems which are embarrassingly parallel, so
> that's not as good of an example.
> 

OK, granted. I just think that what makes the Beowulf concept really
interesting is the fact that with the availability of cheap and
reasonably fast networking technology, such as Myrinet or even good
fast ethernet switches, some truly parallel applications now scale
quite well on clusters of PC's (cf. the results in your own posting
from yesterday). This is where Beowulfs are breaking new ground --
real parallel supercomputing with COTS components -- and why they
provide a cost effective alternative to the "big-iron" MPP machines
like Cray T3E, SGI O2K and IBM SP2 for medium sized problems.

Embarrassingly parallel applications are not as challenging (or
interesting in my humble opinion) since they can run just as well on a
pile of fast workstations connected with 28K modems ;-) Also, people
have been spreading this kind of jobs across networks of workstations
for years and years, so it's not a new idea. Still, I acknowledge that
the Beowulf movement has helped spreading this idea to new areas,
where this type of computing would traditionally have been done on a
large central server. This is for example the case in the group of
solar physicists I work with.

> 
> I haven't. And I've been doing CFD for most of the time. Just goes to
> show that computational science is pretty varied, eh?
> 

Yep, that's why it is so much fun!

Best,
  Rasmus