[mvapich-discuss] problem with intra-node communications

[Christian Guggenberger]

On Thu, May 07, 2009 at 06:01:48PM +0600, Maya Khaliullina wrote:
> Hello,
> When we run any MPI program on 4, 6 or 8 processes using intra-node
> communications only, job hangs at end stage (we believe it occurs during
> MPI_Finalize). But if we use mvapich2-0.98 or Intel MPI 3.1.26 it works
> fine.
> Have any idea about this problem?

Just curious, will your application finish properly if you disable the SRQ feature
at runtime ? e.g. something like

mpiexec -n 4 -env MV2_USE_SRQ 0 ./a.out

I see a quite similar behaviour with MT25208 based SDR cards, but not
(with the very same software environment) with MT25418 [ConnectX IB DDR, PCIe 2.0 2.5GT/s]
based cards.

cheers.
 - Christian