Skip to content



VASP requires a paid license. If you wish to use VASP on the cluster and your research group has purchased a license, please contact us to gain access to the cluster installation. Thank you for your cooperation.

VASP and Slurm

In Slurm, there is big difference between --ntasks and --cpus-per-task which is explained in our Requesting Resources documentation.

For the purposes of VASP, --ntasks-per-node should always equal NCORE (in your INCAR file). Then --nodes should be equal to the total number of cores you want, divided by --ntasks-per-node.

VASP has two parameters for controlling processor layouts, NCORE and NPAR, but you only need to set one of them. If you set NCORE, you don’t need to set NPAR. Instead VASP will automatically set NPAR.

In your mpirun line, you should specify the number of MPI tasks as:

mpirun -n $SLURM_NTASKS vasp_std

Cores Layout Examples

If you want 40 cores (2 nodes and 20 cpus per node):

in your submission script:

#SBATCH --nodes=2
#SBATCH --ntasks-per-node=20
mpirun -n 2 vasp_std



You may however find that the wait time to get 20 cores on two nodes can be very long since cores request via --cpus-per-task can’t span multiple nodes. Instead you might want to try breaking it up into smaller chunks. Therefore, try:

in your submission script:

#SBATCH --nodes=4
#SBATCH --ntasks-per-node=10
mpirun -n 4 vasp_std



which would likely spread over 4 nodes using 10 cores each and spend less time in the queue.

Grace mpi partition

On Grace's mpi parttion, since cores are assigned as whole 24-core nodes, NCORE should always be equal to 24 and then you can just request ntasks in multiples of 24.

in your submission script:

#SBATCH --ntasks=48 # some multiple of 24
mpirun -n $SLURM_NTASKS vasp_std



Additional Performance

Some users have found that if they actually assign 2 MPI tasks per node (rather than 1), they see even better performance because the MPI tasks doesn't span the two sockets on the node. To try this, set NCORE to half of your nodes' core count and increase mpirun -n to twice the number of nodes you requested.

Additional Reading

Here is some documentation on how to optimally configure NCORE and NPAR:

Last update: May 26, 2022