MPIProf is a lightweight, profile-based application performance analysis tool that works with many MPI implementations, including HPE MPT. The tool gathers statistics in a counting mode via PMPI, the MPI standard profiling interface.

This tool is available for either the NAS x86_64 or aarch64 systems.

MPIProf can report profiling information about:

• Collective and/or point-to-point MPI functions called by an application (time spent, number of calls, and message size)
• MPI and POSIX I/O statistics
• Memory used by processes on each node
• Memory and power usage by applications
• Call path information of MPI calls (requires the -g compiler option and the -cpath mpiprof option)

In addition, mpiprof incorporates the mbind process binding capabilities through its various options. For more information about the options, read mpiprof -h after loading the scicon/mpiprof module.

There are two ways you can use MPIProf:

• The mpiprof profiling tool on the command line:

  mpiexec -n <N> mpiprof [-options] [-h|-help] a.out [args]

• MPIProf API routines

mpiprof Usage Example

Because the command-line method does not require changing or recompiling your application, we recommend this method for general use. To run mpiprof, load the modulefiles for your compiler, MPI library, and the mpiprof tool, then run the command line. For example:

module load comp-intel mpi-hpe/mpt
module load scicon/mpiprof
mpiexec -n 128 mpiprof -o mpiprof.out a.out

Note: If the -o option is not included, the profiling results are written to the mpiprof_stats.out file, or the <a.out>_mpiprof_stats.out file if the executable name is known.

For more details about using this tool, see the MPIProf user guide on a NAS system:

/u/scicon/tools/opt/mpiprof/default/doc/mpiprof_userguide.pdf

MPIProf was developed by NAS staff member Henry Jin.