Gromacs 2023.2/2023.3
Several versions of the http://www.gromacs.org molecular dynamics program are installed. The most complete and latest is version 2023.2. Only the single precision version is installed at this time. The various computers are compared with the GMX50 bare water benchmark https://ftp.gromacs.org/pub/benchmarks/ . As shown below, the GPU version is much faster. Because of poor scaling over multiple GPUs, Gromacs is not approved on multiple GPU public-use servers.
These results reflect four separate installations that are autoselected by the module at runtime. Multiple node runs are not yet tested and will require a machinefile and possibly -genv as shown in mpi. These versions are compiled with OpenMP but OMP_NUM_THREADS>1 on CPU-only nodes did not improve results. gromacs/2023.3 may also be selected in both CPU and GPU. Performance appears similar. CPU version with gcc or GPU version with icx have not been compiled
Node Type GPU NS/day partition 32-core dual Intel 6130 none 3.021 comp 32-core dual Intel 6130 1xV100 16.242 gpu 64-core dual AMD 7543 none 7.057 acomp 64-core dual AMD 7543 1xA100 24.775 agpu With $NC=32 or 64 physical cores: CPU: module load intel/23.0.0 python/intel-3.7.7 impi/21.8.0 mkl/22.1.0 gromacs/2023.2 OMP_NUM_THREADS=1 mpirun -np $NC gmx_mpi mdrun -resethway -npme 0 -notunepme -noconfout -nsteps 1000 -v -s bench.tpr Single GPU: sudo nvidia-smi -c 0 #set shared mode# module load gcc/11.2.1 cuda/12.1 impi/19.0.9 python/intel-3.7.7 gromacs/2023.2 OMP_NUM_THREADS=$NC gmx_mpi mdrun -nb gpu -resethway -npme 0 -notunepme -noconfout -nsteps 1000 -v -s bench.tpr
