GROMACS Installations on Tetralith & Sigma

GROMACS is a versatile package to perform molecular dynamics, i.e. simulate the Newtonian equations of motion for systems with hundreds to millions of particles. It is primarily designed for biochemical molecules like proteins, lipids and nucleic acids.

The GROMACS installations on Tetralith and Sigma are compiled with support for MPI+OpenMP execution and are intended to be launched with the NSC mpprun launch wrapper. Some installations have been patched and compiled with PLUMED support, and have corresponding modules tagged with the "PLUMED" label.

How to run

To see which versions are installed use module avail

module avail gromacs
/software/sse/modules:
... snip ...
GROMACS/2018.1-PLUMED-nsc2-gcc-2018a-eb
GROMACS/2018.1-nsc2-gcc-2018a-eb
GROMACS/2018.4-PLUMED-nsc1-gcc-7.3.0-bare
GROMACS/2018.4-nsc1-gcc-7.3.0-bare
GROMACS/2019.2-nsc1-gcc-7.3.0-bare
GROMACS/2019.6-nsc1-gcc-7.3.0-bare

Load the GROMACS module corresponding to the version you want to use. For instance

module load GROMACS/2019.6-nsc1-gcc-7.3.0-bare

Versions older than 5.0.4 will not be covered in these instructions since they deviate significantly and are very old by now.

Example batch script

A minimum batch script for running a standard GROMACS mdrun looks something like this:

#!/bin/bash
#SBATCH -n 128
#SBATCH -t 4:00:00
#SBATCH -J jobname
#SBATCH -A SNIC-xxx-yyy

# Start with a clean environment
module purge

# Substitute with your required GMX module
module add GROMACS/2019.6-nsc1-gcc-7.3.0-bare

mpprun gmx_mpi mdrun <additional GMX job specification options>

Note that you should edit the jobname, account number, number of tasks and requested walltime to your liking before submitting. This should be the default way of running GROMACS mdrun at NSC, and only if you know you need to use OpenMP in addition to MPI (or instead of as the case may be) should you consider hybrid MPI+OpenMP runs.

If you need to run the hybrid, MPI+OpenMP version of GROMACS you should in addition specify to SLURM how many CPU cores should be used per rank using the -c option to SLURM. Match this value when launching gmx_mpi using the -ntomp option and set the OMP_NUM_THREADS in accord. Also request GROMACS to pin ranks and threads with the -pin on option. An example batch script using four Tetralith nodes as above for this is

#!/bin/bash
#SBATCH -n 64
#SBATCH -c 2
#SBATCH -t 4:00:00
#SBATCH -J jobname
#SBATCH -A SNIC-xxx-yyy

# Start with a clean environment
module purge

# Substitute with your required GMX module
module add GROMACS/2019.6-nsc1-gcc-7.3.0-bare

export OMP_NUM_THREADS=${SLURM_CPUS_PER_TASK}
mpprun --pass="-cpus-per-rank ${SLURM_CPUS_PER_TASK}" gmx_mpi mdrun \
    -ntomp ${SLURM_CPUS_PER_TASK} -pin on \
    <additional GMX job specification options>

For most normal uses of GROMACS, running GROMACS in hybrid mode is (lightly) detrimental to performance, which is not to say there aren't use cases for hybrid runs. For performance reasons, it is recommended to keep the number of CPU cores per rank as low as possible and avoid using more than 8, even though this is certainly possible (getting successively worse performance, most likely).

Running a PLUMED-patched GROMACS using PLUMED input may require a great deal more care getting reasonable performance from the run. It is advised to check that your pinning options doesn't oversubscribe the hardware resources. Log in to the nodes of the job (jobsh nXXX) and check pinning with hwloc-ps -t and resource utilisation.

GPU-enabled GROMACS installations at NSC

The GPU-enabled GROMACS installations at NSC are currently thread-MPI only, and can thus only be run on a single GPU-equipped node on Tetralith (or Sigma). The reason for this is that the scalability, i.e. performance increase when adding more nodes/GPUs, of GPU-enabled GROMACS on Tetralith using standard MPI is excessively poor or non-existent because the interconnect is not set up as GPUDirect capable (for technical reasons).

The GPU-enabled GROMACS installations can be distinguished by their being built with the "gcccuda" toolchain, and you can list them with module -t avail gcccuda | grep GROMACS. Since they are tMPI-parallel, the binary is called gmx and it need not be launched with mpprun.

Building GROMACS at NSC

If you for some reason need to compile GROMACS yourself, here are some very basic build instructions which should give you a working, well-performing installation which can be launched with the mpprun launch utility:

module load buildenv-gcc/7.3.0-bare CMake/3.15.2
mkdir temp_gmx_build_dir && cd temp_gmx_build_dir

cmake \
-D CMAKE_C_COMPILER=mpicc \
-D CMAKE_CXX_COMPILER=mpicxx \
-D GMX_OPENMP=ON \
-D GMX_MPI=ON \
-D GMX_BUILD_OWN_FFTW=ON \
-D GMX_BUILD_UNITTESTS=ON \
-D BUILD_TESTING=ON \
/path/to/unpacked/gromacs/source

make -j4
make tests # Builds the tests
make test # Runs the tests

After you have built and tested, you can verify that your build can be launched with mpprun with the dumptag utility. For instance

$ dumptag bin/gmx_mpi 

File name   /some/path/to/bin/gmx_mpi
NSC tags ---------------------------------
Build date      200317
Build time      122236
Built with MPI      openmpi 3_1_2__gcc__7_3_0__bare
Linked with     gcc 7_3_0__bare
Tag version     6
------------------------------------------

If dumptag does not report (or reports wrongly) which MPI it was built with, something has gone wrong with the build, and mpprun will not be able to launch it correctly. With respect to performance, compilers and MPI: GROMACS has so far performed best using GCC and OpenMPI on Tetralith, and dramatic changes in this regard are not expected for various reasons. Stick to these build tools is our recommendation.

If you are interested in how NSC builds and tests our GROMACS installations, see e.g. the file /software/sse/manual/GROMACS/2019.6/g73/nsc1/build.txt on Tetralith/Sigma.