I've had a few off-list questions about our Infiniband support in
LAM/MPI 7.1.
Short version: LAM requires that the IB stack support something called
"shared completion queues."
Longer version:
It looks like we left out some magic IB keywords in the LAM v7.1
release notes. We used a feature of the IB drivers called "shared
completion queues" which did not appear in the Mellanox VAPI stack
until their version 3.0. We tested and developed the LAM IB RPI with
Mellanox VAPI v3.2.
Some other IB stacks still do not yet support shared completion queues,
and therefore will not work with LAM/MPI's IB RPI module. The behavior
that you may see is LAM hanging for anything more than 2 processes
using the IB RPI. Or it may crash altogether. Neat feature, eh? :-)
The end result is that you want to check and see if your IB stack
supports shared completion queues.
Hope that helps clarify some things.
--
{+} Jeff Squyres
{+} jsquyres_at_[hidden]
{+} http://www.lam-mpi.org/
|
