[torqueusers] Intel mpiexec support

David Golden dgolden at cp.dias.ie
Thu Oct 27 08:12:06 MDT 2005


On 2005-10-27 15:38:06 +0200, Jacques Foury wrote:
> Paul Van Allsburg a écrit :
> 
> >I was using Intel's MPI.  That's going to make it a little more 
> >difficult to setup and run parallel jobs within Torque.
> 
> we're using Intel Compilers (C/C++, Fortran) with MPICH2. We just had to 
> configure/compile mpicc, mpif90... using the Intel compilers, and we're 
> calling mpiexec within the Torque's batches.
>

N.B. Intel have an MPI implementation, too. [1]

Note again that there is confusion over mpiexec: There is an "mpiexec" from
in OSC that interfaces mpich and various relatives to pbs and various
relatives (e.g. torque).  [2]

But mpi-2 standardised the "mpiexec" name as the name for
executables that kick off mpi processes (see "too many mpiexecs"
at [2]), so some mpi-2 implementations now
come with executables called  "mpiexec" that are no relation to the 
OSC mpiexec[-for-pbs].

Some MPI implementations with "native" support for the PBS TM API,
e.g. LAM, don't need the OSC mpiexec[-for-pbs] glue. 

I've never actually used Intel's MPI, but it's apparently based
on mpich2 [3]. According to [2], mpiexec[-for-pbs] mpich2 support is also
compatible with the Intel MPI, but you might want to scan through
the mpiexec[-for-pbs] mailing list for issues with it and patches.

Worst comes  to worst, though, you can presumably kick off 
whatever it needs based on the contents of the file
pointed to by $PBS_NODEFILE. But that can lead to
inaccurate accounting, and depressingly often stray processes that
need to be reaped.

[1] http://www.intel.com/cd/software/products/asmo-na/eng/cluster/mpi/index.htm
[2] http://www.osc.edu/~pw/mpiexec/
[3]http://www.intel.com/cd/software/products/asmo-na/eng/cluster/mpi/219891.htm


More information about the torqueusers mailing list