Page Comparison

GROMACS (GROningen MAchine for Chemical Simulations) is a molecular dynamics simulation package.

...

updated: May 2023

...

...

 Available version

Note:

[a] GROMACS/2023.2-cpeGNU-23.09-CUDA-12.0 MPI does NOT have PME GPU decomposition feature. This does not affect the normal usage of GROMACS unless you have very large system (i.e. >10 M atoms), please see Massively Improved Multi-node NVIDIA GPU Scalability with GROMACS | NVIDIA Technical Blog for more details about PME GPU decomposition feature.

1. Input file

The input file of GROMACS mdrun command is TPR file (.tpr). For an example TPR files, you can see https://www.mpinat.mpg.de/grubmueller/bench where GROMACS intensive benchmark sets are provided.

2. Job submission script

create a script using vi submit.sh command and specify the following details depending on computational resources you want to use.

2.1 using compute node (1 node)

#!/bin/bash
#SBATCH -p compute      	#specific partition
#SBATCH -N 1 -c 128         #specific number of nodes and taskcores per nodetask
#SBATCH -t 5-00:00:00       #job time limit <hr:min:sec>
#SBATCH -A lt999999         #project name
#SBATCH -J GROMACS      	#job name

##Module Load##
module purgerestore
module load GROMACS/2022.5-GNU-11.2-CUDA-11.7

gmx mdrun -deffnm input 

The script above using compute partition (-p compute), 1 node (-N 1) with 128 cores per task (-c 128) for 1 task (default). The wall-time limit is set to 5 days (-t 5-00:00:00) which is the maximum. The account is set to lt999999 (-A lt999999) that is subjected to change to your own account. The job name is set to GROMACS (-J GROMACS ).

2.2 using compute node (>1 node)

#!/bin/bash
#SBATCH -p compute      		         #specific partition
#SBATCH -N 4 --ntasks-per-node=64 -c 2 	 #specific number of nodes, task per node, and taskcores per nodetask
#SBATCH -t 5-00:00:00        		     #job time limit <hr:min:sec>
#SBATCH -A lt999999             		     #project name
#SBATCH -J GROMACS      		         #job name

##Module Load##
module purgerestore
module load GROMACS/2022.5-GNU-11.2-CUDA-11.7

srun -c $SLURM_CPUS_PER_TASK gmx_mpi mdrun -deffnm input -ntomp 2$SLURM_CPUS_PER_TASK

The script above using compute partition (-p compute), 4 node (-N 4) with 2 cores per task (-c 2) for 64 tasks per node (--ntasks-per-node=64). This result in 2 x 64= 128 cores per node and 4 x 128= 512 cores in total. The wall-time limit is set to 5 days (-t 5-00:00:00) which is the maximum. The account is set to lt999999 (-A lt999999) that is subjected to change to your own account. The job name is set to GROMACS (-J GROMACS ).

...

2.3 using GPU node (1

...

card)

#!/bin/bash
#SBATCH -p gpu                          #specific partition
#SBATCH -N 1 --ntasks-per-node=1 -c 16  #specific number of nodes, task per node, and taskcores per nodetask
#SBATCH --gpus-per-task=1		        #specific number of GPU per task
#SBATCH -t 5-00:00:00                   #job time limit <hr:min:sec>
#SBATCH -A lt999999                       #project name
#SBATCH -J GROMACS                   	#job name

##Module Load##
module purgerestore
module load GROMACS/2022.5-GNU-11.2-CUDA-11.7

gmx mdrun -deffnm input -update gpu

The script above using gpu partition (-p gpu), 1 node (-N 1) with 16 cores per task (-c 16) and 1 GPU card per task (--gpus-per-task=1) for 1 tasks per node (--ntasks-per-node=1). This result in 1 x 16= 16 cores with 1 x 1= 1 GPU card. The wall-time limit is set to 5 days (-t 5-00:00:00) which is the maximum. The account is set to lt999999 (-A lt999999) that is subjected to change to your own account. The job name is set to GROMACS (-J GROMACS ).

2.4 using GPU node (>1 cards)

#!/bin/bash
#SBATCH -p gpu                          	#specific partition
#SBATCH -N 1 --ntasks-per-node=4 -c 16  	#specific number of nodes, task per node, and taskcores per nodetask
#SBATCH --gpus-per-task=1		            #specific number of GPU per task
#SBATCH -t 5-00:00:00                     	#job time limit <hr:min:sec>
#SBATCH -A lt999999                       	#project name
#SBATCH -J GROMACS                   	    #job name

##Module Load##
module purgerestore
module load GROMACS/2022.5-GNU-11.2-CUDA-11.7

export#export GMX_GPU_DD_COMMS=true
export#export GMX_GPU_PME_PP_COMMS=true
export GMX_ENABLE_DIRECT_GPU_COMM=true

gmx mdrun -deffnm stmvinput -update gpu -nb gpu -bonded gpu -pme gpu -ntmpi 8 -ntomp 8 -npme 1

The script above using gpu partition (-p gpu), 1 node (-N 1) with 16 cores per task (-c 16) and 1 GPU card per task (--gpus-per-task=1) for 1 tasks per node (--ntasks-per-node=4). This result in 4 x 16= 64 cores with 4 x 1= 4 GPU card. The wall-time limit is set to 5 days (-t 5-00:00:00) which is the maximum. The account is set to lt999999 (-A lt999999) that is subjected to change to your own account. The job name is set to GROMACS (-J GROMACS ).

Note: Two environment variables (line 13, 14) are set to enable GPU direct communication. see Massively Improved Multi-node NVIDIA GPU Scalability with GROMACS | NVIDIA Technical Blog for more detail.

In GROMACS 2022 and 2023, the GMX_GPU_DD_COMMS and GMX_GPU_PME_PP_COMMS are removed, please use GMX_ENABLE_DIRECT_GPU_COMM instead, see Environment Variables — GROMACS 2022 documentation and Environment Variables - GROMACS 2023 documentation for detail.

3. Job submission

 using sbatch submit.sh command to submit the job to the queuing system.

...