[mpich-discuss] MPI on a cluster of dual-CPU machines

[Pavan Balaji]

> What is the most efficient way of doing this ... which options should
> I specify while configuring the software and how do I specify the
> options to mpdboot and mpiexec to use this feature of MPI ?

Just use the default configuration in the latest release of MPICH2; 
it'll automatically figure out that there are multiple CPUs on the same 
node. For mpdboot, there are different ways of specifying this:

$ cat hostfile
host1
host2

... this means that the first process will be on host1, second on host2, 
then the third process will go back to host1, the fourth on host2, etc.

Another way is:

$ cat hostfile
host1:2
host2:2

... this means that the first two processes will be on host1, and the 
next two on host2.

If you are using the second way, make sure you use the --ncpus option to 
mpdboot.

  $ mpdboot -f hostfile -n 2 --ncpus=2

  $ mpiexec -n 4 ./app

Also remember that the "-n" option in mpdboot specifies the number of 
physical nodes, and not the number of CPUs, while in mpiexec, it refers 
to the number of processes.

If you don't want to deal with all these things, an alternate approach 
is to bypass mpd and use the new Hydra process manager. You can just do:

$ mpiexec.hydra -f hostfile -n 4 ./app

... you won't need mpdboot in this case.

Hope that helps.

  -- Pavan

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji