On Mon, Jul 28, 2008 at 12:08 PM, Terry Dontje <Terry.Dontje@sun.com>
Is it something to do with the MPI_Barrier call? osu_latency uses MPI_Barrier and from rhc's email it sounds like his code does too.
Jeff Squyres wrote:
On Jul 28, 2008, at 12:03 PM, George Bosilca wrote:
Interesting. The self is only used for local communications. I don't expect that any benchmark execute such communications, but apparently I was wrong. Please let me know the failing test, I will take a look this evening.
FWIW, my manual tests of a simplistic "ring" program work for all combinations (openib, openib+self, openib+self+sm). Shrug.
But for OSU latency, I found that openib, openib+sm work, but openib+sm+self hangs (same results whether the 2 procs are on the same node or different nodes). There is no self communication in osu_latency, so something else must be going on.
I don't think it's an issue with MPI_Barrier(). I'm running into this problem with srtest.c (one of the example programs from the mpich distribution). It's a ring-type test with no barriers until the end, yet it hangs on the very first Send/Recv pair from rank0 to rank1.
I my case, openib and openib+sm works, but openib+self & openib+sm+self hang.