mpiexec and mpich/p4 shared memory [was Re: [PBS-USERS] mpiexec with OpenPBS_2.3.16]

[Anupam Sharma]

Hi Pete,

	The second option works for my linux PII cluster.  I was searching
on the web and I found one more person having similar problem as me with
SMP machines.  I reconfigured mpich with comm=shared enabled but that
didn't help...in fact I had more problems just with mpirun doing that.

	However, the second option works for me so I am glad. Thanks for
your help.

anupam

----------------- your mail ----------------

| pw at osc.edu said:
| > The mpiexec code is designed to use shared memory communication when you
| > run more than one process on a node.  I could actually add an option to
| > disable this, but I just assumed people would always want it.  I'll
| > think about that and let you know if I make that work.
|
| If you really do not want to recompile your mpich to add
| --with-comm=shared, you can at least now configure mpiexec to
| startup mpich/p4 codes without shared memory on SMP nodes.
|
| Grab the latest from the cvs and configure as such:
|
|     cvs -d :pserver:anoncvs at cvs.osc.edu:/cvs co mpiexec
|     cd mpiexec
|     ./configure --with-default-comm=p4 --disable-p4-shmem
|
| I chose not to make it an mpiexec run-time option, although that would
| be easy, since this is not something that users should turn on and off at
| will.  It is absolutely imperative that mpich/p4 and mpiexec agree on
| this shared memory setting.  The way jobs are started is quite different
| in the two cases.
|
| In other words, do one of the following scenarios but don't mix them:
|
|     cd /usr/local/mpich
|     ./configure --with-comm=shared ..
|     cd /usr/local/mpiexec
|     ./configure ..
|
| or
|
|     cd /usr/local/mpich
|     ./configure ..
|     cd /usr/local/mpiexec
|     ./configure --disable-p4-shmem ..
|
| I recommend that you try the first case if you can.
|
| 		-- Pete
|