CASTEP¶

For full details on the on this software package please refer to the CASTEP documentation.

Building CASTEP¶

The first requirement is a copy of the .tar file containing the CASTEP libraries and installer. Extracting the .tar file will create the CASTEP-[version] directory.

tar -xf  CASTEP-[version].tar

Place the file linux_x86_64_ifort19.mk in the directory CASTEP-[version]/obj/platforms. From within the CASTEP-[version] directory run the following command:

make COMMS_ARCH=mpi FFTW=fftw3 MATHLIBS=mkl10

and follow the install instructions when prompted (the instructions are set in the makefile).

To make the executable accessible to Allinea/ARM MAP add the option BUILD=intermediate to the make command:

make COMMS_ARCH=mpi FFTW=fftw3 MATHLIBS=mkl10 BUILD=intermediate

In the following we will make use of two trial datasets, provided on the CASTEP website: TiN, Al3x3, and DNA.

Preparing a CASTEP job¶

Basic information about the content of the input files and way CASTEP will execute the job can be found by doing a dryrun of the job first: castep.serial [job-name] -dryrun

Here the castep implementation is serial, as the castep.mpi command does not appear to be able to perform a dryrun

Note

In order to perform a dryrun on the nextgenio system it is necessary to have the serial implementation of castep available. This can be done conducting the initial build of the application that uses the make command specified above, and follow this with the command make install. This may result in the Error running make install during CASTEP build, however the castep.serial executable should be created nonetheless.

Various parameters for the run can be set in the [job-name].param file. Possibly relevant parameters (from the perspective of performance) are max_scf_cycles, the number of iterations over the grid, and write_checkpoint, which writes the full analysis to file in a time consuming step. The values can be set by adding a line to the .param file, or by editing the relevant line if the option is already included.

The verbosity of the output files ( [job-name].castep ) can also be set in the [job-name].param file with the iprint option, which takes values 0-3 in increasing level of verbosity. For maximum verbosity add the following line to the .param file :

iprint : 3

The recommended minimum number of cores to request for a CASTEP run is the total number of k-points (representing the sampling grid of electron wave vectors). The k-points for a specific run setup can be found by doing a dryrun first. Per k-point the calculation is spread further (to solve for the plane waves), and split among the available threads.

Executing a CASTEP job¶

The following is an example batch script to submit to the job scheduler that will run the analysis of the TiN dataset:

#!/bin/bash

#SBATCH --nodes=1
#SBATCH --ntasks=8
#SBATCH --cpus-per-task=6

#SBATCH -D /path/to/TiN-directory
#SBATCH -o /path/to/TiN-directory/TiN.out.%A.%N.log
#SBATCH -e /path/to/TiN-directory/TiN.err.%A.%N.log
#SBATCH --job-name=castep-tin

export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK

srun "/path/to/castep-build/CASTEP-18.1/obj/linux_x86_64_ifort19/castep.mpi" "path/to/TiN-directory/TiN-mp"

For the TiN, Al3x3, and DNA test sets, in the standard configuration, the number of k-points is 8, 2, and 1 respectively. In the example above one MPI process is assigned per k-point, and the plane wave calculations are spread over six cpus, each running one thread.

Performance of the system will depend both on the parallelisation of the run and on the memory mode selected.

The type of parallelisation can be controlled in the batch script using the --ntasks option and the OMP_NUM_THREADS variable. Some effects of different parallelisation are explored in Performance Analysis Tools.

Initial results indicate that running CASTEP on a node, or nodes, in App Direct mode has a considerable performance benefit over a run in Memory mode.