On Apr 18, 2007, at 8:44 AM, stephen mulcahy wrote:
> ~/openmpi-1.2/bin/mpirun --mca btl_tcp_if_include eth0 --mca btl
> tcp,self --bynode -np 2 --hostfile ~/openmpi.hosts.80
> ~/IMB/IMB-MPI1-openmpi -npmin 2 pingpong
> Neither one resulted in a significantly different benchmark.
That's truly odd -- I can't imagine why you wouldn't get the TCP
transport with the above command line. But the latencies, as you
mentioned, are far too low for TCP.
To be absolutely certain that you're not getting the IB transport, go
to the $prefix/lib/openmpi directory and make a bogus subdirectory
(e.g., "mkdir foo"). Then move the *openib* and *mvapi* components
into that subdirectory (you'll likely only have either "openib" *or*
"mvapi" -- not both).
Then try running again.
What this does is move the plugins out of the directory where Open
MPI will find them, so it will be forced to use TCP. If this works,
it would seem to indicate that there's a problem somehow that the
remote nodes are ignoring the specification to use tcp,self, which
would be surprising (but could be a bug).