AMBER

Examples at https://github.com/hpc-cofc/example-runs/tree/master/chemistry/AMBER

AMBER 18

Idea

This calculation simulates a 136475-atom biological molecule for 100,000 time steps (0.2 nanoseconds). It also reports how long of a simulation we can run in a day if we were to run it in

• one CPU-only node (one-node) with 40 cores = 40 MPI processes

• two CPU-only nodes (two-nodes) with 80 cores in total = 80 MPI processes

• one node with NVIDIA Quadro P4000 GPU (gpu-quadro-p4000) = 1 process on CPUs with massive processing on GPU

• one node with NVIDIA Tesla V100 GPU (gpu-tesla-v100) = 1 process on CPUs with massive processing on GPU

It demonstrates

• the power parallelization using MPI

  • compare calculations on a single node (smp) vs. multiple nodes (mpi)

• the power of GPUs

  • compare calculations on CPUs (smp, mpi) vs. GPUs (p4000, v100)

• different between consumer-grade and datacenter GPUs

  • compare consumer-grade (NVidia Quadro P4000) and datacenter (NVIDIA Tesla V100) GPU

How to run

On the CofC HPC cluster, you can go into the one-node, two-nodes, gpu-quadro-p4000 and gpu-tesla-v100 directories and enter sbatch run.slurm

Results

Here is a summary from a sample run.

Calculation	Number of MPI processes	GPUs?	Speed of simulation (ns/day) [higher is better]
./single-node/40.out:	40	-	9.17
./two-nodes/80.out:	80	-	14.49
./gpu-quadro-p4000/quadro.out:	1	1	31.45
./gpu-tesla-v100/24.out:	1	1	116.84