To run a hybrid MPI/OpenMP* program, follow these steps:
Make sure the thread-safe (debug or release, as desired) Intel® MPI Library configuration is enabled (release is the default version). To switch to such a configuration, source mpivars.[c]sh with the appropriate argument, see Selecting Library Configuration for details. For example:
$ source mpivars.sh release
Set the I_MPI_PIN_DOMAIN environment variable to specify the desired process pinning scheme. The recommended value is omp:
$ export I_MPI_PIN_DOMAIN=omp
This sets the process pinning domain size to be equal to OMP_NUM_THREADS. Therefore, if for example OMP_NUM_THREADS is equal to 4, each MPI process can create up to four threads within the corresponding domain (set of logical processors). If OMP_NUM_THREADS is not set, each node is treated as a separate domain, which allows as many threads per MPI process as there are cores.
For pinning OpenMP* threads within the domain, use the Intel® compiler KMP_AFFINITY environment variable. See the Intel compiler documentation for more details.
Run your hybrid program as a regular MPI program. You can set the OMP_NUM_THREADS and I_MPI_PIN_DOMAIN variables directly in the launch command. For example:
$ mpirun -n 4 -genv OMP_NUM_THREADS=4 -genv I_MPI_PIN_DOMAIN=omp ./myprog
Intel® MPI Library Developer Reference, section Tuning Reference > Process Pinning > Interoperability with OpenMP*.