HPC/Applications/lammps/Package OMP

Package OMP

LAMMPS modules since 2012 are compiled with yes-user-omp, permitting multi-threaded runs of selected pair styles, and in particular MPI/OpenMP hybrid parallel runs. To set up such runs, see HPC/Submitting and Managing Jobs/Advanced node selection.

Be careful how to allocate CPU cores on compute nodes.

• The number of MPI tasks running on a node is determined by options to mpirun.
• The number of threads that each MPI task runs with is usually determined by the environment variable OMP_NUM_THREADS, which is 1 by default on Carbon.
• The LAMMPS package omp command has a mandatory argument Nthreads to either override OMP_NUM_THREADS, or use it when set to * - the recommended practice. It appears to be sensible to specify a concrete thread number only once, which means in the job file, and let LAMMPS inherit this value with Nthreads set to *.

Usage

1. Use the command package omp Nthreads mode near the beginning of your LAMMPS control script.
2. Do one of the following:

   • Append /omp to the style name (e.g. pair_style lj/cut/omp)
   • Use the suffix omp command.
   • On the command line, use the -suffix omp switch.

3. In the job file or qsub command line, reserve nodes and ppn suitable for OpenMP runs.
4. Call the lmp_openmpi (regular) binary.

Input file example

Examples:

• Recommended:

package omp * force/neigh

• Not recommended:

package omp 4 force

Job file example

• Single node - choose from:
  • pure OpenMP, single entire node, or
  • pure OpenMP, single node, possibly shared.

And for the main command:

lmp_openmpi -suffix omp -in infile

• Multiple nodes - use:
  • HPC/Submitting and Managing Jobs/Advanced node selection#OpenMP/MPI hybrid.

And for the main command:

mpirun … lmp_openmpi -suffix omp -in infile