VASP on LUMI¶
VASP (Vienna Ab Initio Simulation Package) is a popular software package for "atomic scale materials modelling from first principles" (see the VASP home page). In general, it runs well on LUMI-C.
There is currently no version of VASP that can use the AMD GPUs in the Early Access Platform or LUMI-G.
We provide automatic installation scripts for several versions of VASP. The procedure in general is described on the EasyBuild page. The step by step procedure to install VASP 6.3.0 is:
- Download the VASP source code "vasp.6.3.0.tgz" from the VASP portal.
- Upload the file somewhere to your home directory on LUMI.
- Load the LUMI software environment:
module load LUMI/21.12.
- Select the LUMI-C partition:
module load partition/C.
- Load the EasyBuild module:
module load EasyBuild-user.
Then, you can run the install command:
eb --sourcepath=[directory where the VASP source is stored] VASP-6.3.0-cpeGNU-21.12.eb -r
Afterwards, you will have a module called "VASP/6.3.0-cpeGNU-21.12" installed in your home directory.
module load VASP/6.3.0-cpeGNU-21.12
The usual VASP binaries,
vasp_gam etc will now be in your PATH. Launch VASP with e.g.
srun vasp_std. Please note that you must do
module load LUMI/21.12 partition/C to see the VASP module in the module system. The same applies to the SLURM batch scripts which you send to the compute nodes.
You can see other versions of VASP that can be automatically installed by running the EasyBuild command
eb -S VASP
or checking the LUMI-EasyBuild-contrib repository on GitHub directly.
Example batch scripts¶
A typical batch job using 4 compute nodes and MPI only:
#!/bin/bash #SBATCH -J GaAs512 #SBATCH -N 4 #SBATCH --partition=small #SBATCH -t 00:30:00 #SBATCH --mem=200G #SBATCH --exclusive --no-requeue #SBATCH -A project_XYZ #SBATCH --ntasks-per-node=128 #SBATCH -c 1 export OMP_NUM_THREADS=1 ulimit -s unlimited module load LUMI/21.12 partition/C VASP/6.3.0-cpeGNU-21.12 srun vasp_std
A typical batch job with MPI and 8 OpenMP threads per rank:
#!/bin/bash #SBATCH -J GaAs512 #SBATCH -N 4 #SBATCH --partition=small #SBATCH -t 00:30:00 #SBATCH --mem=200G #SBATCH --exclusive --no-requeue #SBATCH -A project_XYZ #SBATCH --ntasks-per-node=16 #SBATCH -c 8 export OMP_NUM_THREADS=8 export OMP_PLACES=cores export OMP_PROC_BIND=close ulimit -s unlimited module load LUMI/21.12 partition/C VASP/6.3.0-cpeGNU-21.12 srun vasp_std
Due to the restrictive license conditions imposed by the VASP group, which forbids sharing of any kind of benchmark numbers with anyone (including other licensed users!), we get cannot give detailed recommendations on the best way to run VASP.
In general, try increasing NCORE to 16-32, and NSIM to 32. OpenMP works best with many threads, eg. OMP_NUM_THREADS=8. It is best to run with all 128 procressor cores per compute node if you can, but reducing the number of cores per compute node does not decrease performance as much as you might expect. That can be useful when you are constrained by memory and need more available memory per MPI rank. If possible, use k-point parallelization up to the maxium number of k-points.