Okay, I fixed this in r24536. Sorry for the problem, Damien - thanks for catching it! Went unnoticed because the folks at the Labs always use IB.
On Mar 16, 2011, at 7:20 PM, Ralph Castain wrote: > I believe I see the problem - and why it wouldn't show up for IB. It looks > like the hier module passes an incorrect flag to the modex unpack function, > which causes that function to place the modex values as attributes assigned > to the node instead of a process, rather than placing the values into the > modex database. So when you look up a value, you get a single value for the > entire node. > > Works for IB because the interface info is at the node level. Doesn't work > for TCP because the "interface" info is at the proc level. > > Since it was only tested on IB before, this didn't show up. Should be easy to > fix. > > On Mar 16, 2011, at 6:15 PM, Jeff Squyres wrote: > >> On Mar 16, 2011, at 5:37 PM, George Bosilca wrote: >> >>> I just checked and IB does work correctly. But then I remembered that IB is >>> different, the connection are peer based, so they don't happens during the >>> modex exchange. The data is exchanged over RML messages, but outside the >>> modex. >> >> Not quite. The openib BTL does use the modex to send around connection >> information. The actual connections are made lazily -- just like the TCP >> BTL -- but the OOB CPC (i.e., the default connection mode in the openib BTL) >> uses RML to do the 2/3 way handshake. That's all. >> >> But the point here is: the openib BTL does rely on the modex. >> >> -- >> Jeff Squyres >> jsquy...@cisco.com >> For corporate legal information go to: >> http://www.cisco.com/web/about/doing_business/legal/cri/ >> >> >> _______________________________________________ >> devel mailing list >> de...@open-mpi.org >> http://www.open-mpi.org/mailman/listinfo.cgi/devel >
