I'm usually working on 64 bits nodes and I never had this issue. I
just run your test program on 64 bits AMD (compiled with gcc with and
without debug), using MX, mvapi, openib and tcp and the code just run
without any problems. There is one thing I see it can make your code
failing. If you have a multi-devices cluster (namely 2 ethernet
cards), it might happen that Open MPI does not discover them
correctly. Or at least does not configure/use them correctly. There
is an entry in the FAQ about how to do that. If it's the case then
forcing the usage of one specific interface will solve the problem.
On Apr 5, 2006, at 2:24 PM, Charles Shuller wrote:
> I encountered an error with send/recv pairs on AMD64. The first send
> completes successfully, but subsequent send/recv pairs hang. I also
> believe that MPI_Bcast also hangs under simmilar conditions, but have
> not created an explicit testcase for this (it happened in the code I
> was working on). Both machines are running AMD64 processors.
> The problem only manifests when the tasks are on seperate machines.
> The problem does not occur on x86 (32-bit) machines.
> Charles Shuller
> users mailing list