[Beowulf] Update on mpi problem

Dries Kimpe Dries.Kimpe at cs.kuleuven.be
Sat Jul 12 08:04:36 PDT 2008

* Ashley Pittman <apittman at concurrent-thinking.com> [2008-07-10 12:44:26]:

> That doesn't necessarily follow, if you are posing your sends before
> your receives then you are relying on unexpected message buffering
> within the MPI library.  How much of this is available is up the the
> library, not the standard so I think it's possible that openmpi is being
> MPI compliant in both cases.

The latter is easy to check: replace all MPI_Send and MPI_Rsend by
MPI_Ssend. (MPI_Isend / MPI_Irsend -> MPI_Issend)

You can do this through the profiling interface
(create a library that provides MPI_Isend, MPI_Irsend, MPI_Send, MPI_Rsend 
and just calls PMPI_(I)ssend; link it before linking mpi)

If it hangs, it will be at the point where the application relies on
buffering within the MPI library. 


More information about the Beowulf mailing list