Scripts

Scripts

A selection of scripts used for benchmarking VASP on the University of Bath's Balena HPC cluster.

Contents

• CPUTest.py : A script for benchmarking the CPU version of VASP by sweeping the KPAR, NPAR and/or NSIM parameters.

• GPUTest.py : A script for benchmarking the GPU version of VASP by sweeping the number of MPI processes and the NSIM parameter.

• GetTimings.py : A command-line script for quickly extracting timing information from the files in VASP job folders.

• Shared.py : A module containing functions for extracting information from VASP output files; imported by CPUTest.py, GPUTest.py and GetTimings.py.

• PadCSVs.py : Rewrites CSV files to pad rows to a consistent length; required for the "pretty" display on the GitHub website.

Usage

CPUTest.py and GPUTest.py are configured through parameters set at the top of the scripts; these are annotated with explanatory comments and fairly self explanatory.

These scripts are designed to be run from the scheduler - see the example SLURM scripts CPUTest.balena.slm and GPUTest.balena.slm for running on Balena.

GetTimings.py is called from the command line:

python GetTimings.py CPU-GeTe_256/vasp_gam

Analysing "CPU-GeTe_256/vasp_gam"...
  -> # SCF steps: 32
  -> Avgerage t_SCF [s]: 21.93
  -> Elapsed time [s]: 745.55
  -> Final E_0 [eV]: -670.88396000

The optional --skip_scf_cycles=N argument can be used to exclude the first N SCF steps from the average cycle time.

PadCSV.py is a utility script, again called from the command line; type PadCSV.py -h for usage instructions.

N.B. The default python on Balena does not have the argparse module imported by GetTimings.py - use the distribution provided by the python/2.7.8 module instead.

Acknowledgements

The CUDA setup in GPUTest.balena.slm is based on the notes from Peter Larsson's blog post, Running VASP on Nvidia GPUs.