[Beowulf] Multidimensional FFTs

Greg Lindahl lindahl at pathscale.com
Thu Mar 2 13:53:42 PST 2006


On Thu, Mar 02, 2006 at 12:32:40PM -0800, Konstantin Kudin wrote:

>  By the way, any idea why gigE on Opterons with Broadcom chokes so
> easily with different flavours of "alltoall" operation, such as if done
> via "isend+irecv"? If functioning flawlessly, the scaling would have
> been not too bad, but there seem to be issues of some sort ...

Congestion. Ethernet switches are the guilty party, not the Broadcom
nics. MPI can makes this better by avoiding, for example, having
everyone send first to process 0, then to 1, etc. So MPICH, for
example, has rank R send to R+1, R+2, etc.

-- greg




More information about the Beowulf mailing list