map_rank_to_gpu

#!/bin/bash
#
# Copyright (C) 2002-2021 the Network-Based Computing Laboratory (NBCL), The
# Ohio State University.
# 
# Contact: Dr. D. K. Panda (panda@cse.ohio-state.edu)
# 
# For detailed copyright and licensing information, please refer to the
# copyright file COPYRIGHT in the top level directory.
# 
# Name        : map_rank_to_gpu
# 
# Description : This script is used to set the affinity between an MPI process
# and the GPU on the system.
#
#               The processes rank on a node is normally used to do this and
#               different MPI launchers expose this information through
#               different environment variables. For example, MVAPICH2 uses
#               MV2_COMM_WORLD_LOCAL_RANK and OpenMPI uses
#               OMPI_COMM_WORLD_LOCAL_RANK. For other MPI libraries, please set
#               the LOCAL_RANK variable below to the appropriate one.
#
#               Some python applications don't respect CUDA device mappings.
#               It's much more convenient to force CUDA_VISIBLE_DEVICES here
#               rather than setting it within python, because
#               CUDA_VISIBLE_DEVICES must be set after MPI_Init but before the
#               offending python package is imported. 
#
#               NOTE: Only use map_rank_to_gpu if process mapping is not
#               respected. Forcing CUDA_VISIBLE_DEVICES will disable CUDA IPC,
#               which may impact performance

export LOCAL_RANK=$MV2_COMM_WORLD_LOCAL_RANK 
export CUDA_VISIBLE_DEVICES=$MV2_COMM_WORLD_LOCAL_RANK
exec $*