changed a year ago  
Published Linked with GitHub

Running Gromacs at UPPMAX

This page describes how to run the GROMACS molecular dynamics software on UPPMAX systems. See the gromacs web page for more information.

Have a look on this page as well - best practices running GROMAC on HPC.

Selected setups for benchmarking on HPC2N as examples.

Loading the gromac module

$ module load gromacs/2021.1.th

SBATCH script

adapted from HPC2N

#!/bin/bash -l
#SBATCH -A SNIC_project
#SBATCH -t 00:15:00
#SBATCH -p node -n 10
# Use 2 threads per task
#SBATCH -c 2

module load gromacs/2021.1.th

# Automatic selection of single or multi node based GROMACS
if [ $SLURM_JOB_NUM_NODES -gt 1 ]; then
  GMX="gmx_mpi"
  MPIRUN="mpirun"
  ntmpi=""
else
  GMX="gmx"
  MPIRUN=""
  ntmpi="-ntmpi $SLURM_NTASKS"
fi

# Automatic selection of ntomp argument based on "-c" argument to sbatch
if [ -n "$SLURM_CPUS_PER_TASK" ]; then
  ntomp="$SLURM_CPUS_PER_TASK"
else
  ntomp="1"
fi
# Make sure to set OMP_NUM_THREADS equal to the value used for ntomp
# to avoid complaints from GROMACS
export OMP_NUM_THREADS=$ntomp
$MPIRUN $GMX mdrun $ntmpi -ntomp $ntomp -s MEM.tpr -nsteps 10000 -resethway

How important is to select appropriate options

Here is a simple benchmark ran on single interactive node with 20CPUs using the MEM example from this benchmark https://www.mpibpc.mpg.de/grubmueller/bench.

$ module load gromacs/2021.1.th
$ mpirun -np XX gmx_mpi mdrun -ntomp YY -s MEM.tpr -nsteps 10000 -resethway

where XX * YY = 20



























$ grep "gmx_mpi\|MPI ranks\|Performance" *

#md.log.1#:  gmx_mpi mdrun -ntomp 1 -s MEM.tpr -nsteps 10000 -resethway
#md.log.1#:On 12 MPI ranks doing PP, and
#md.log.1#:on 8 MPI ranks doing PME
#md.log.1#:Performance:       20.520        1.170

#md.log.2#:  gmx_mpi mdrun -ntomp 2 -s MEM.tpr -nsteps 10000 -resethway
#md.log.2#:On 10 MPI ranks, each using 2 OpenMP threads
#md.log.2#:Performance:       25.037        0.959

#md.log.3#:  gmx_mpi mdrun -ntomp 4 -s MEM.tpr -nsteps 10000 -resethway
#md.log.3#:On 5 MPI ranks, each using 4 OpenMP threads
#md.log.3#:Performance:        5.388        4.454

#md.log.4#:  gmx_mpi mdrun -ntomp 5 -s MEM.tpr -nsteps 10000 -resethway
#md.log.4#:On 4 MPI ranks, each using 5 OpenMP threads
#md.log.4#:Performance:       24.090        0.996

#md.log.5#:  gmx_mpi mdrun -ntomp 10 -s MEM.tpr -nsteps 10000 -resethway
#md.log.5#:NOTE: Your choice of number of MPI ranks and amount of resources results in using 10 OpenMP threads per rank, which is most likely inefficient. The optimum is usually between 1 and 6 threads per rank.
#md.log.5#:On 2 MPI ranks, each using 10 OpenMP threads
#md.log.5#:Performance:        3.649        6.577

md.log:  gmx_mpi mdrun -ntomp 20 -s MEM.tpr -nsteps 10000 -resethway
md.log:Performance:        2.012       11.931

Notice how bad is the last run
$ mpirun -np 1 gmx_mpi mdrun -ntomp 20 -s MEM.tpr -nsteps 10000 -resethway (lines 25-26)

According to this short test, this particular setup runs best on single Rackham node with
$ mpirun -np 10 gmx_mpi mdrun -ntomp 2 -s MEM.tpr -nsteps 10000 -resethway (lines 8-10)

Running older versions of gromacs

Versions 4.5.1 to 5.0.4:

The gromacs tools have been compiled serially. The mdrun program has also been compiled in parallel using MPI. The name of the parallel binary is mdrun_mpi.

Run the parallelized program using:

mpirun -np XXX mdrun_mpi

… where XXX is the number of cores to run the program on.

Version 5.1.1

The binary is gmx_mpi and (e.g.) the mdrun command is issued like this:

mpirun -np XXX gmx_mpi mdrun

Contacts:


tags: UPPMAX, SNIC
Select a repo