more on structural models for clusters
James.P.Lux at jpl.nasa.gov
Thu Oct 2 20:39:33 EDT 2003
At 05:29 PM 10/2/2003 -0700, Bill Broadley wrote:
>On Wed, Oct 01, 2003 at 03:36:35PM -0700, Jim Lux wrote:
> > In regards to my recent post looking for cluster implementations for
> > structural dynamic models, I would like to add that I'm interested in
> > "highly distributed" solutions where the computational load for each
> > processor is very, very low, as opposed to fairly conventional (and widely
> > available) schemes for replacing the Cray with a N-node cluster.
> > The number of processors would be comparable to the number of structural
> > nodes (to a first order of magnitude)
>Er, why bother? Is there some reason to distribute those things so
>thinly? Your average dell can do 1-4 Billion floating point ops/sec,
>why bother with so few per CPU? Am I missing something?
Your average Dell isn't suited to inclusion as a MCU core in an ASIC at
each node and would cost more than $10/node... I'm looking at Z80/6502/low
end DSP kinds of computational capability in a mesh containing, say,
Sure, we'd do algorithm development on a bigger machine, but in the end
game, you're looking at zillions of fairly stupid nodes. The commodity
cluster aspect would only be in the development stages, and because it's
much more likely that someone has solved the problem for a Beowulf (which
is fairly loosely coupled and coarse grained) than for a big multiprocessor
with tight coupling like a Cray.
Haven't fully defined the required performance yet, but, as a starting
point, I'd need to "solve the system" in something like 100
microseconds. The key is that I need an algorithm for which the workload
scales roughly linearly as a function of the number of nodes, because the
computational power available also scales as the number of loads.
Clearly, I'm not going to do a brute force inversion or LU decomposition of
a 100,000x100,000 matrix... However, inverting 100,000 matrices, each,
say, 10x10, is reasonable.
James Lux, P.E.
Spacecraft Telecommunications Section
Jet Propulsion Laboratory, Mail Stop 161-213
4800 Oak Grove Drive
Pasadena CA 91109
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
More information about the Beowulf