GROMACS and PLUMED with GPU

Since version 4.6.x GROMACS can run in an hybrid mode making use of both your CPU and your GPU (either using CUDA or OpenCL for newer versions of GROMACS). The calculation of the short-range non-bonded interactions is performed on the GPU while long-range and bonded interactions are at the same time calculated on the CPU. By varying the cut-off for short-range interactions GROMACS can optimize the balance between GPU/CPU loading and obtain amazing performances.

GROMACS patched with PLUMED takes into account PLUMED in its load-balancing, adding the PLUMED timings to the one resulting from bonded interactions and long- range interactions. This means that the CPU/GPU balance will be optimized automatically to take into account PLUMED!

It is important to notice that the optimal setup to use GROMACS alone on the GPU or GROMACS + PLUMED can be different, try to change the number of MPI/OpenMP processes (OpenMP) used by GROMACS and PLUMED to find optimal performances. Remember that in GROMACS multiple MPI threads can use the same GPU:

i.e. if you have 4 cores and 2 GPU you can:

• use 2 MPI/2GPU/2OPENMP:

export PLUMED_NUM_THREADS=2
mpiexec -np 2 gmx_mpi mdrun -nb gpu -ntomp 2 -pin on -gpu_id 01

• use 4 MPI/2GPU:

export PLUMED_NUM_THREADS=1
mpiexec -np 4 gmx_mpi mdrun -nb gpu -ntomp 1 -pin on -gpu_id 0011

Of notice that since plumed 2.5 and gromacs 2018.3 the number of openMP threads can automatically set by gromacs (so PLUMED_NUM_THREADS is not needed, and the number of OpenMP threads used by plumed is set by -ntomp)

mpiexec -np 2 gmx_mpi mdrun -nb gpu -ntomp 2 -pin on -gpu_id 01