large clusters and topologies

[Greg Lindahl]

> - I suppose Ethernet is not all too efficient with this
> size, even with switches instead of hubs. What's the largest
> practical size for an Ethernet 'wulf?

Depending on the details, Ethernet can work great for larger machines. The
1,000 cpu genetic algorithm machine (www.genetic-programming.com) uses
ethernet, Google's 5,000 cpu cluster uses ethernet, Incyte's 3,000 cpu
cluster uses ethernet, and so forth. The key is that their application's
network bandwidth and latency demands aren't maxing out what they have.

> Any alternatives to Myrinet upwards of that?

Not really. It depends.

> - Which type of algorithm can take direct advantage of a cube,
> toroid, ... topology? I would think even a "dumb" job like
> matrix multiplication might benefit from a 2D lattice (instead
> of the typical all-to-all unique segment). Comments?

There are lots of books which discuss this topic. There have been many
religious wars fought over the subject.

-- greg