User Tools

Site Tools


gromacs

This is an old revision of the document!


Gromacs

Several versions of the http://www.gromacs.org molecular dynamics program are installed. The most complete and latest is version 2016.3.

Gromacs makes heavy use of the http://www.fftw.org FFT package. Both Gromacs and FFTW make heavy use of Intel SSE/AVX vector instructions. Our modules for gromacs 2016.3 and fftw 3.3.6 automatically select the proper vector type for our compute nodes at runtime (SSE2, AVX, AVX2 for FFTW, same plus SSE4.1 for Gromacs). The scheduler requires that multiple nodes in a job are of the same type, so the vector type selected for the head compute node should be correct for all.

Non-GPU usage

with single/double precision selected by _mpi or _mpi_d:

module purge
module load intel/16.0.1 impi/5.1.2 mkl/16.0.1 fftw/3.3.6  gromacs/2016.3
NP=$(wc -l <$PBS_NODEFILE)
mpirun -np $NP -machinefile $PBS_NODEFILE {mdrun_mpi|mdrun_mpi_d} -c conf.gro -s topol.tpr

Benchmark results (single node) http://www.gromacs.org/GPU_acceleration ADH Cubic PME

Node Type                Vector SP ns/day DP ns/day
12-core Intel 2-X5670    SSE4.1     3.16      6.47
16-core Intel 2-E5-2670     AVX     7.47     12.45
24-core Intel 2-E5-2650v4  AVX2    10.95     16.78       
32-core AMD 4-Opteron 6136 SSE2     4.93      8.55

In many of our benchmarks, a single 32-core 6136 Trestles node is about the same performance as a 16-core E5-2670 node, but AMD does not work very well for Gromacs because of its limited SSE capability.

GPU usage

in progress

gromacs.1494441802.txt.gz · Last modified: 2017/05/10 18:43 by root