mpiexec, intel, and topspin

[Pete Wyckoff]

roskar at astro.washington.edu wrote on Wed, 27 Jun 2007 02:05 -0700:
> I am trying to install mpiexec on our local mini-supercomputer - we  
> use the topspin IB, intel compilers and torque/moab.
> 
> I've tried installing using the simple procedure outlined in the  
> README file. I've tried running the test program, but nothing happens  
> (well, the string of '.' keeps growing larger). I've tried searching  
> the web for some info, and two things have become apparent:
> 
> - how do I need to configure mpiexec to use the topspin libraries, I  
> simply specified the ib device to be the default?
> 
> - do I need to compile differently if we want to use code compiled  
> with intel compilers?

No need to have IB libraries around when compiling mpiexec.  Nor
do you need a separate executable for topspin or intel.  You can
change the type of MPI at runtime.

However, to run the test program ("hello"), you need to compile that
like any other MPI application.  So your own PATH should have
"mpicc" in it, and that should be the correct version.  And at
runtime, you may need LD_LIBRARY_PATH.  Standard stuff here.

The growing line of dots may indicate that qsub never got the job to
run.  Type "qstat" in another window and see.

To avoid the PBS issues, get yourself an interactive job and test
mpiexec there:

    qsub -I -l nodes=3 -l walltime=1:00:00

(or similar).  In the shell it gives you, try to build the hello
prorgram "make hello" and run the new mpiexec "./mpiexec hello".

		-- Pete