the best slurm script for the lammps with DPD on heterogeneous architecture

LAMMPS LAMMPS Mailing List Mirror
1

Dear Lammps users,

I'm moving to a new cluster with GPUs so I wonder what would be the
most productive option for the run on it. This cluster has 8 CPUs per
node and one GPU. So is it possible to use all CPUs if I use GPU
package (because I need pair dpd and newer gpgpu packages does not
support it)?
In case if I run with the slurm script like the following:
#SBATCH --nodes=1
#SBATCH --cpus-per-task=1
#SBATCH --ntasks-per-node=8

I get the error: "Accelerator sharing is not currently supported on system".

2

Dear Lammps users,

I'm moving to a new cluster with GPUs so I wonder what would be the
most productive option for the run on it. This cluster has 8 CPUs per
node and one GPU. So is it possible to use all CPUs if I use GPU
package (because I need pair dpd and newer gpgpu packages does not
support it)?

you can always compile in the GPU package, even if you don't use it.
however, you will not be using the GPU, if you don't use any GPU
accelerated pair styles, so it is a pointless exercise. however, the
GPU package *does* support dpd, so i don't understand your statement,
why it would not support it. how do you know?

In case if I run with the slurm script like the following:
#SBATCH --nodes=1
#SBATCH --cpus-per-task=1
#SBATCH --ntasks-per-node=8

I get the error: "Accelerator sharing is not currently supported on system".

that is a message from the batch system. you have to discuss this with
the system manager(s). this has nothing to do with LAMMPS.

axel

3

Dear Lammps users,

I'm moving to a new cluster with GPUs so I wonder what would be the
most productive option for the run on it. This cluster has 8 CPUs per
node and one GPU. So is it possible to use all CPUs if I use GPU
package (because I need pair dpd and newer gpgpu packages does not
support it)?

you can always compile in the GPU package, even if you don't use it.
however, you will not be using the GPU, if you don't use any GPU
accelerated pair styles, so it is a pointless exercise. however, the
GPU package *does* support dpd, so i don't understand your statement,
why it would not support it. how do you know?

Yes, GPU package support dpd, I meant that new packages like KOKKOS do
not have dpd, so I use GPU package.

In case if I run with the slurm script like the following:
#SBATCH --nodes=1
#SBATCH --cpus-per-task=1
#SBATCH --ntasks-per-node=8

I get the error: "Accelerator sharing is not currently supported on system".

that is a message from the batch system. you have to discuss this with
the system manager(s). this has nothing to do with LAMMPS.

I see, thanks.