Re: Approximating Hypothetical Beowulfs

Eugene Leitl (eugene.leitl@lrz.uni-muenchen.de)
Fri, 4 Jun 1999 18:13:26 -0700 (PDT)

Dan Clemmensen writes:
> I'm fairly sure (gut feel, not mathematical analysis) that you
> can make much more cost-effective use of a 4-NIC solution by
> establishing a four-dimensional hypergrid than by bonding.
>
> Of course, it depends on the problem.

Yes. There are problems which scale bad with the numbers of node, and hence require few nodes with a better networks. One would want to use a hypernet (all nodes interconnected) or a channel bonded network in this case.

Then there are embarrassingly parallel problems, which could utilize a torus, or a diamond lattice network topology (the latter handy for 3d physical problems) which could scale to thousands of node, in some cases linearly. Of course, one can also combine the multiple-NIC and multiple-switch approach, to create both channel-bonded few-node clusters and extremely large 3d lattice clusters using switches as networking glue. Of course component failure will preclude building a cluster from significantly more than 1000 nodes. Let's wait for systems on a chip....