Re: MPI/RT for Myrinet

On Fri, Jul 29, 2005 at 02:45:44PM +0530, Ashish Bijlani wrote:

> But MPICH-GM itself doesn't support deterministic communication

Note that basically no network provides deterministic communication if
there is an error. And congestion is the kind of problem that no
current MPI interconnect does it in a way that leads to deterministic
results. You can probably put such a protocol on top of Myrinet, but
it's a lot of work.

So it really depends on what you're doing: how hard of a guarantee
would you like?

-- greg