Single-node Run

Now we are going to load the package we just installed. Since only one version is installed, we can plainly load the gromacs@2021.1 module.

spack load gromacs@2021.1

To be explicit we can use the hash we kept form the install phase.

spack load /${GROMACS_101_HASH}

Now let’s get going. We are running on a 72 vCPU instance, we will start 18 MPI ranks with 4 OpenMPI threads each. This will decompose the benchmark system (benchRIB) into 18 domains, each domain will be computed using 4 threads.

A first iteration towards an optimal performance - we’ll dive deeper later in the workshop.

mkdir -p /fsx/jobs/101
cd /fsx/jobs/101
mpirun -n 18 gmx_mpi mdrun -ntomp 4 -s /fsx/input/gromacs/benchRIB.tpr -resethway

This will kick of the GROMACS run with 10000 steps, reset half way to smooth the benchmarking results.

The job run should take around 5 minutes. At the end it will sumarize the performance.

# execution spec instance Ranks x Threads ns/day
1 native gromacs@2021.1 c5n.18xl 18 x 4 4.7

Afterwards, please log out of the instance and remove the allocation.