On Thu, May 07, 1998 at 03:59:39PM -0400, Alexander Kushnirenko wrote:
> > I've seen several responses point you to Beowulf. Check it out.
> I have a look. It's much more than we want or capable of doing. You
> build supercomputer out of small computers connected with fast network, so
> that speed of calculation is proportional to the number of Linux boxes.
Almost proportional. What some people don't get (or don't want to) is the
the speed of the network is a very limiting factor. If you want to build a
Beowulf cluster on top of El Cheapo NE-2000 10 Mbps, it won't work. But
doing the same using Good Ethernet 100 Mbps achieves quite a lot... I can
dig my bookmarks if you are interested.
The other limiting factor is homogeneousness <sp>. On top of a homogeneous
MPI cluster, the thing performs quite well. On top of *extremely*
heterogeneous PVM cluster performance suffers a LOT, unless of course you
design the programs with this in mind, but that's an entirely different
issue. Hands on experience beats any book or reference here.
> They succeded in solving problems like fluid motion and multi-body
> gravitational calculations, and speed is indeed proportional to number of
> linux boxes, not quite linearly though. It would be probably too ambitious
> for us.
It depends on the scale of problems you are after. 100**3 grids on fluid
dynamics (quite small) take any available PC down to its knees. 50x50
lattices (which can prove useful for certain kind of research) can be
handled reasonably fast on most modern PC's, being a little pattient.
Being potentially intrusive, at Fermi *there's* a Beowulf cluster running.
Cheers,
Marcelo
--
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
