profit.run.slurm

Scheduling runs on a HPC cluster with SLURM

• targeted towards aCluster@tugraz.at

• each run is submitted as a job using a slurm batch script

• run arrays are submitted as a slurm job array

• by default completed runs are recognised by the interface, but the scheduler is polled as a fallback (less often)

Module Contents

Classes

SlurmRunner

Runner which submits each run as a job to the SLURM scheduler on a cluster

class profit.run.slurm.SlurmRunner(*, interface='zeromq', cpus=1, openmp=False, custom=False, path='slurm.bash', options=None, command='srun profit-worker', **kwargs)

Bases: profit.run.runner.Runner

Runner which submits each run as a job to the SLURM scheduler on a cluster

• generates a slurm batch script with the given configuration

• can also be used with a custom script

• supports OpenMP

• tries to minimize overhead by using job arrays

• polls the scheduler only at longer intervals

property config

__repr__()

Return repr(self).

spawn(params=None, wait=False)

spawn a single run

Parameters:

• params – a mapping which defines input parameters to be set

• wait – whether to wait for the run to complete

spawn_array(params_array, wait=False, progress=False)

spawn an array of runs

maximum ‘parallel’ at the same time blocking until all are submitted

poll(run_id)

check the status of the run directly

poll_all()

cancel(run_id)

cancel_all()

clean()

remove generated scripts and any slurm-stdout-files which match slurm-*.out

generate_script()