verifying SMP

Rusty Lusk lusk at
Fri Jul 26 11:57:05 PDT 2002

| Is there a good reference for configure options to mpich btw? 

Goodness is in the eye of the beholder, of course, but what you have to work
with is the output of

  configure --help

and the User's Guide for the device that you configure with, in the mpich/doc

Plus you have the address mpi-bugs at for questions and problems.

| for example, I am under the impression that the "-with-comm=shared" option
| optimizes the ch_p4 device for clusters of SMP nodes; however, I have been
| unable to verify this.

What --with-comm=shared does for the ch_p4 device is add the capability for
messages to be transmitted through shared memory with memcpy when the source
and destination processes are on the same SMP (and TCP when not).  Having to
check both sources for incoming messages causes (instead of just relying on
select) introduces some overhead, and some find that their applications do not
perform much better than when they just use TCP for all communication.  You
will need to try it on your application to make sure.  Multi-method
communication will be handled a little better in our new implementation of

| Anyway, where we stand now is that it is beyond the capabilities of mpich to 
| bind a process to a processor.  

This is true, but need not be the case forever.  Suggestions on just exactly
how you think this should work are welcome.  

Rusty Lusk

More information about the Beowulf mailing list