MPI+OMP in AOCC

export OMP_NUM_THREADS=2
export OMP_PLACES=cores
export OMP_PROC_BIND=close

mpirun -np 48 \
  --map-by ppr:12:numa:pe=2 \
  --bind-to core \
  ./lmp -sf omp -pk omp 2 -in in.AK1

 

FlagMeaning

-np 48	Use 48 MPI ranks total
--map-by ppr:12:numa:pe=2	12 ranks per NUMA node × 4 = 48 ranks; each gets 2 cores (for OMP)
--bind-to core	Binds each rank and its threads to specific physical cores
OMP_PLACES=cores	Each thread sticks to its allocated core
OMP_PROC_BIND=close	Threads are kept near their parent MPI rank

 

 

 

Updated in 250421

export OMP_NUM_THREADS=2
export OMP_PROC_BIND=spread
export OMP_PLACES=cores

mpirun -np 32 --map-by numa:PE=2 --bind-to core \
        -x OMP_NUM_THREADS \
        -x OMP_PLACES \
        -x OMP_PROC_BIND \
        lmp -sf omp -in benchmark.in