XICSRT on Multiple Processors¶

XICSRT has built-in support for raytracing over multiple processors through the use of Python’s multiprocessing library. To use this functionality one only needs to replace the call xicsrt.raytrace() with xicsrt.raytrace_mp().

Windows Support¶

Multiprocessing on Windows requires that xicsrt.raytrace_mp() is wrapped in a name == "__main__" test. In a python script or jupyter notebook on Windows this means replacing the call to xicsrt.raytrace() with the following:

if __name__ == "__main__":
    results = xicsrt.raytrace_mp(config)

The Command Line Interface can be used without any modifications.

Cluster Computing¶

A command line interface to xicsrt is available to enable computations on a computer cluster.

A single call to xicsrt can utilize multiple processors but is (currently) limited to run on a single computational node. However, it is easy to combine results from multiple calls to xicsrt allowing multiple nodes to be used. The use of slurm Job Arrays is recommended.

To launch a single call to xicsrt on 16 processors using slurm the following command can be used:

srun -n1 -c16 xicsrt config.json --mp --numruns 16 --processes 16

For multiple parallel calls to xicsrt use the --suffix option to give all output files a unique name.

Note

For multiple calls to xicsrt make sure that the random seed is either

equal to None (the default)
set to a different value for each call using the --seed argument.

Job Array Example

Below is a simple example of a slurm batch file to run xicsrt on 64 processors over 4 nodes.

job.sh

#!/bin/bash
#SBATCH -J xicsrt
#SBATCH -o ./job_%A_%a.out
#SBATCH -e ./job_%A_%a.err
#SBATCH --nodes=1
#SBATCH --cpus-per-task=16
#SBATCH --array=0-3

srun xicsrt config.json --mp --numruns 16 --processes 16 --suffix $SLURM_ARRAY_TASK_ID &> xicsrt.log

To send this job to the queue type sbatch job.sh at the command line.

After the jobs are complete the saved results can be read using xicsrt_io.load_results() and combined using xicsrt_raytrace.combine_raytrace().