Lammps simulations

LAMMPS LAMMPS General Discussion
1

Hi everyone,
I am running an MD simulation in LAMMPS for a polymer system consisting of approximately 20,000 atoms and the OPLS-AA force field. However, the performance seems quite slow — I’m currently achieving only about 4 ns/day.
Here are the job submission details:
#SBATCH -n 10
#SBATCH --gres=gpu:1
#SBATCH --mem-per-cpu=2048
I’m using this command to run the simulation:
lmp -sf gpu -in run.in.nvt
I would really appreciate any input on the following:

  1. what is the expected time simulations per day?
  2. Any suggestions for improving the performace?
  3. Best practices for running lammps simulations on GPU.

Thanks in Advance.

2

Impossible to say. Specifically without knowing anything about your hardware and the details of your system geometry and simulation settings.

Having 10 nodes with 1 GPU each leaves you with 2000 atoms per GPU. That is far too few to reach good utilization. Your system is too small. You may get better performance using CPUs across that many nodes.

There is a discussion the manual: 7. Accelerate performance — LAMMPS documentation

Generally, you should do a proper scaling test to determine the optimal performance settings.