multiprocessing - mpirun gets higher rank than requested with --ntasks from sbatch script -
i launching mpi program through sbatch script (corresponding example script provided system administrators):
#! /bin/bash -l #sbatch --job-name=test #sbatch -o stdout.log #sbatch -e stderr.log #sbatch --ntasks=160 #sbatch --time=0-00:10:00 #sbatch --qos=normal cd $work/ mpirun ./mpiprogram
however, seems more mpi processes launched ntasks. resulting in 200 processes when 160 requested, different numbers. note nodes have 16 or 20 cores. of worker processes (all pretty identical in case) run slower others, perhaps because of swapping. swapping may caused many processes on 1 node, causing them use memory.
should specify number of threads mpirun using $slurm_ntasks
? or going on here?
Comments
Post a Comment