Humboldt-Universität zu Berlin - Mathematisch-Naturwissen­schaft­liche Fakultät - Institut für Physik

OpenMPI - Message Passing Interface

Version:

OpenMPI-1.8.4

Architectures:

x86_64

Compile commands:

  C C++ F77 F90
GNU mpicc mpiCC mpif77 mpif90
Intel mpicc-icc mpiCC-icpc mpif77-ifort mpif90-ifort
Portland mpicc-pgcc mpiCC-pgCC mpif77-pgf77 mpif90-pgf90
  • Make sure that /usr/global/mpi/bin is in PATH.
  • Compilation with -static is not supported.
  • For the GNU and Intel compilers, the f77 and f90 scripts invoke the same compiler (gfortran, ifort) with different switches.
  • The naming of mpiCC-icpc reflects the correct invocation of Intel's C++ compiler: icpc.
  • Compilers used to build OpenMPI:
    GNU 4.7.2 (Debian Wheezy)
    Intel 15.0.1
    Portland 14.6

Run a parallel program

Start it from the command line:

Make sure that /usr/global/mpi/bin is in PATH.

      mpirun -np 4 a.out      # start 4 processes on the local machine
      mpirun -host node1,node2,node3,node4 a.out      # run on the named nodes
      mpirun -hostfile <filename> a.out         # run on nodes listed in hostfile

Start it from a PBS (torque) batch script:

The node information given to the batch job is sufficient. OpenMPI automatically imports the nodelist from PBS (torque). Don't specify the nodes with the run command again, just say

      mpirun a.out

Notes:

  • Don't mix architectures (i386/amd64) in a parallel run.
  • Note the differences to MPICH.

Documentation:

First of all, there are manpages for commands and MPI routines.
See also http://www.open-mpi.org/doc/v1.6/.
In addition, some commands accept the switch -h to display usage notes.

License:

free

letzte Änderung: B Bunk, 30.06.2015