Gromacs on ThetaGPU
What is Gromacs?
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 that have a lot of complicated bonded interactions, but since GROMACS is extremely fast at calculating the nonbonded interactions (that usually dominate simulations) many groups are also using it for research on non-biological systems, e.g. polymers.
Using GROMACS at ALCF
ALCF offers assistance with building binaries and compiling instructions for GROMACS. For questions, contact us at email@example.com.
- Download latest source code: http://manual.gromacs.org/documentation/2022.1/download.html
tar -xzf gromacs-2022.1.tar.gz
- Submit an interactive job to a ThetaGPU compute node from Theta login node:
user@thetalogin4:~>module load cobalt/cobalt-gpu user@thetalogin4:~>qsub -I -n 1 -t 60 -q single-gpu -A PROJECT --attrs filesystems=home Job routed to queue "single-gpu". Wait for job 10108666 to start... Opening interactive session to thetagpu06-gpu0 ... user@thetagpu06:~$
module load cmake
cmake -DCMAKE_C_COMPILER=mpicc -DCMAKE_CXX_COMPILER=mpicxx \ -DBUILD_SHARED_LIBS=OFF -DGMX_BUILD_OWN_FFTW=ON \ -DCMAKE_INSTALL_PREFIX=/path-to/gromacs-2022.1/build \ -DGMX_MPI=ON -DGMX_OPENMP=ON -DGMX_GPU=CUDA \ -DCUDA_TOOLKIT_ROOT_DIR=/user/local/cuda-11.4
make –j 16
- The installed binary is
Running Gromacs on ThetaGPU
Prebuilt Gromacs binaries can be found in the directory
A sample qsub script follows that will run GROMACS on a full node using all eight GPUs available.
#!/bin/bash -l #COBALT -n 1 #COBALT -t 30 #COBALT -q full-node #COBALT -project catalyst #COBALT --attrs filesystems=home,theta-fs0 NODES=`cat $COBALT_NODEFILE | wc -l` mpirun -hostfile $COBALT_NODEFILE --np 8 \ /soft/applications/gromacs/gromacs_cuda/gmx_mpi.2022.1 \ mdrun -ntomp 8 -gputasks 01234567 -nb gpu -pme gpu -npme 1 \ -dlb yes -resethway -pin on -v deffnm step5_1 -g test.log
We strongly suggest that users try combinations of different numbers of nodes, MPI ranks per node, number of GPU tasks/devices, GPU task decomposition between nonbonded and PME kernels, and OMP threads per rank to find the optimal throughput for their particular workload.
The following is a representative benchmark for a system with 30,000 atoms generated on a single ThetaGPU node with above example.
|Core time(sec)||Wall time(sec)||(%)|