[Beowulf] Clos networking and myrinet

[Greg Lindahl]

On Wed, Jun 08, 2005 at 03:10:10PM -0400, Mack.Joseph at epamail.epa.gov wrote:

> o If all nodes in a beowulf were exchanging data in pairs
> before a barrier and then later were exchanging data in
> a different combination of pairs before another barrier,
> which pairwise exchanges would be blocked?

None.

> o Is a Clos network ever non-blocked?

Yes, but only for that trivial case. If you have a 2D computation and
you want to exchange with 4 neighbors, it's usually fastest to do all
irecvs, all isends, and then wait. You don't want to divide it up and
try 4 groups of synchronized stuff to avoid blocking. That ends up
being slow.

> o Why does myrinet use Clos switching for beowulfs, if
> block/non-blocking isn't an issue?

Because a Clos network is a good network, not because it's formally
non-blocking in a certain situation. Or at least that's my opinion, I
can't speak for Danny or Chuck.

-- greg