Skip to content

NAMD GPU Resident Mode #158

@michaelmarty

Description

@michaelmarty

Hi TACC Team,
I am a new user of TACC and was setting up NAMD3 to run in GPU resident mode. When I tried the method in the docs, it gave errors. It might be worth adding a new section on GPU resident setup, and I wanted to report what worked for me.

I didn't try non-resident methods, which I think are what is used in the docs. All this was on LS6. I used the A100, but I'm testing it on H100 as well. Should work on A100-small, but I haven't tested it there yet.

To set up GPUresident mode, you need "GPUresident on" inserted in the input file.

Then, I set it up using this SLURM file:

#SBATCH -N 1            # Total number of nodes
#SBATCH -n 1            # Total number of mpi tasks
#SBATCH -p gpu-a100         # Queue name
#SBATCH -t 24:00:00     # Run time (hh:mm:ss) - 24 hours

module load gcc/11.2
module load mkl
module load cuda/12.2
module load namd_gpu_resident/3.0.2

cd $SLURM_SUBMIT_DIR

ibrun namd3 +setcpuaffinity step7.1_production.inp > step7.1.log

The performance was really good here, at 100 ns/day for my system. I didn't do extensive benchmarking, but I did explore a few different optimization options, and this seemed to be best. Thought I would pass this on in case it is helpful. Thanks,
Michael Marty

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions