Floating point exception with Intel MPI 2019.x using one task per node

Owned by Marcus Boden (Unlicensed)
Juli 01, 2020
1 min read

Problem

When using Intel MPI 2019 (impi/2019.x) to start MPI jobs executing one task per node, the jobs aborts immediately with an error message

srun: error: gcnXXXX: task 0: Floating point exception
srun: error: gcnYYYY: task 1: Floating point exception
[...]


Solution

This is due to a problem with Intel's Hydra process manager. A workaround is to assign the environment variable I_MPI_HYDRA_TOPOLIB the value ipl (internal default value hwloc).


#!/bin/bash
#SBATCH [...]

export I_MPI_HYDRA_TOPOLIB=ipl

[...]