VASP

VASP#

VASP (Vienna Ab initio Simulation Package) is a first-principles code for electronic-structure calculations and ab initio molecular dynamics using DFT, the projector-augmented-wave (PAW) method, and plane-wave basis sets.

For further information, see the official VASP documentation.

Available versions on Aire#

Version

Load Command

6.5.1

module load vasp/6.5.1/intel-2024.2.1_impi-2021.13_hdf5-1.14.6

6.3.2

module load vasp/6.3.2/intel-2024.2.1_impi-2021.13_hdf5-1.14.6

Executables provided: vasp_std, vasp_gam, vasp_ncl

Note: Versions 6.3.2 and 6.5.1 are currently available for user testing.
Users should validate their results before publication.

Licensing#

Each user must supply their own valid vasp.license file obtained from the VASP portal.

VASP automatically looks for this file in your home directory:

$HOME/vasp.license

If you store it elsewhere, set:

export VASP_LICENSE_FILE=/path/to/vasp.license

The licence file must be readable by your job environment.

Scratch space#

VASP produces large temporary and output files such as WAVECAR, CHGCAR, PROCAR, and vasprun.xml. These files can grow to tens or hundreds of gigabytes depending on system size, so choosing the right scratch location is essential for performance and cleanup.

  • Multi-node jobs: Use the shared NVMe flash filesystem provided by Aire — **$TMP_SHARED** (path: /mnt/flash/tmp/job.<JOB-ID>). This is fast, visible to all nodes in your job, and automatically deleted when the job finishes.

  • Single-node jobs: Use node-local temporary storage — **$TMP_LOCAL** or **$TMPDIR** (path: /tmp/job.<JOB-ID>). This is the fastest option for I/O-intensive runs confined to one node, but it is not shared across nodes.

Best practice

  • Keep only the files required for restart or post-processing (WAVECAR, CHGCAR, OUTCAR).

  • Avoid writing outputs to $HOME. It is slower, backed up, and has strict quotas.

  • Always monitor disk usage; VASP can generate 10–100 GB or more of scratch data for large systems.

How to submit a job#

Select the appropriate binary:

Executable

Purpose

vasp_gam

Γ-point-only calculations (fastest when applicable).

vasp_std

General spin-polarised calculations.

vasp_ncl

Non-collinear magnetism or spin–orbit coupling.

Example job (VASP 6.x – Intel MPI build)#

#!/bin/bash
#SBATCH --job-name=vasp_std
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=64
#SBATCH --time=04:00:00

module purge
module load vasp/6.5.1/intel-2024.2.1_impi-2021.13_hdf5-1.14.6
# (use 6.3.2/... for that version)

# Licence file (if not in $HOME)
#export VASP_LICENSE_FILE=$HOME/vasp.license

# Pure MPI run
export OMP_NUM_THREADS=1

# Run with Slurm + PMIx
srun --mpi=pmix vasp_std > vasp.out 2>&1

Notes:

  • The module provides Intel MPI 2021.13 and automatically configures Slurm bootstrap via I_MPI_HYDRA_BOOTSTRAP=slurm (validated on Aire PMIx v5).

  • Do not load openmpi or other MPI stacks.

  • Aire builds are MPI-only. Keep OMP_NUM_THREADS=1

Common pitfalls#

  • Mixed MPI stacks – never load OpenMPI with the Intel MPI build.

  • Licence not found – ensure $VASP_LICENSE_FILE or $HOME/vasp.license is valid.

  • Scratch space full – write to $TMPDIR or /mnt/scratch/$USER/vasp_scratch.

  • Threading mismatch – leave OMP_NUM_THREADS=1 (pure MPI build).

Performance notes#

Performance characteristics have been validated on Aire compute nodes using the Intel oneAPI 2024.2.1 runtimes and Intel MPI 2021.13 stack.

  • Initial tests show both VASP 6.3.2 and 6.5.1 scale efficiently up to 64 MPI ranks per node.

  • Scaling beyond 64 ranks per node gives diminishing returns; cross-node scaling depends on system size and FFT grid.

  • Optimal configuration: one MPI rank per physical core, OMP_NUM_THREADS=1. (Ref: VASP Wiki: Parallelization)

  • vasp_gam is faster than vasp_std for Γ-point-only calculations.

  • Node-local scratch ($TMPDIR) yields measurable I/O gains vs. shared storage.