SGI's Message Passing Toolkit

The Message Passing Toolkit (MPT) from SGI provides an alternative implementation of MPI. More information is available in the user guide

Modules

The correct environment for using MPT can be set up by loading the mpt module (i.e. module load mpt). The which command can be used to check which version of mpif90 or mpirun will be found (e.g. which mpif90 should return /opt/sgi/mpt/mpt-2.06/bin/mpif90 for MPT).

Building code

If you have loaded the mpt module then the MPT version of the mpif90 command should be found instead of the Intel version. More details on compiling and linking using MPT can be found in Chapter 3 of the MPT user guide

MPT can also work with executables compiled for other types of MPI. See chapter 6 of the user guide for details.

Running on a single node

There is an mpirun command for MPT e.g.

mpirun -np 12 myprog

to run myprog using 12 MPI processes on one node.

If you have problems with mpirun make sure that your path is set up correctly to pick up the MPT version of mpirun and not the Intel MPI version. You may need to load the MPT module in the job file to make sure the correct version is found.

Running multi-node MPT jobs

To run on more than one node you can use the mpiexec_mpt command e.g.

mpiexec_mpt -np 24 myprog

will run myprog using a total of 24 MPI processes on the nodes listed in the PBS nodes file. An example PBS script for running multi-node MPT jobs can be found at /usr/local/examples/qsub_script/example-mpt.pbs.

Performance profiling

Running an MPT code using the perfcatch command will produce profiling information. First load the perfcatcher module using module load perfcatcher then run with

mpirun -np 12 perfcatch myprog

A file called MPI_PROFILING_STATS will be produced which is a text file containing the profiling data.

-- DavidAcreman - 08 Dec 2008