[Beowulf] Re: torus versus (fat) tree topologies

[Greg Lindahl]

On Mon, Nov 15, 2004 at 03:32:15PM -0500, Rick Friedman wrote:

> for CFM, I'm a
> little surprised you appear to care so much about latency, since I'd
> expect your workload to have the usual volume/surface-area scaling, and
> thus doing a lot of work in a single node, and needing only moderate
> bursts of bandwidth for nontrivial problem sizes.

CFD problems come in lots of different kinds, and you seem to only
have run into time-explicit problems where you run with a large grid
until it reaches steady-state. Those have the least amount of
communication.  Other kinds of problems have less data per cpu, and
care about latency. Or you can be using an implicit technique, which
is usually a huge matrix computation, which can be solved in a variety
of ways, some of which are lower-communication than others.

In short, you don't want to ever make any any generalization about CFD
as a whole. You need to specify the solution technique and data size
at a minimum.

-- greg