hpclauncher

Generic interface for submitting jobs to super computer queue managers.

Installation and short introduction

Login to cluster and install hpclauncher as user

python setup.py install --user

Add ~/.local/bin to your PATH

export PATH=~/.local/bin:$PATH

Define a yaml cluster configure file and copy it under

~/.hpclauncher/local_cluster.yaml

For example cluster parameter files see examples/cluster_config. Cluster configure file can also be overriden with HPCLAUNCHERCLUSTER environment variable.

You can now submit jobs from yaml files with

submitYAMLJob.py myjob.yaml

For example job files see examples/job_config.

For testing purposes, adding -t will only print the submission script without submitting it.

submitYAMLJob.py myjob.yaml -t

Jobs can also be created and launched from python:

from hpclauncher import *
# create job. Jobs can be submitted to queue managers
j = BatchJob(jobname='somename', queue='normal',
             nproc=12, timereq=TimeRequest(10, 30, 0), logfile='log_somelog')
# job can contain multiple tasks (commands)
j.append_new_task('echo {message}', message='hello')
submit_jobs(j, testonly=True, verbose=False)

For python examples see examples/python.

List of common keywords

Cluster parameters

• scriptpattern: job script header filled with keyword placeholders '{nprocs}'
• submitexec: executable used to submit jobs
• mpiexec: executable for running parallel jobs, e.g. 'ibrun' or 'mpiexec -n {nproc}'
• useremail: email address where notifications will be sent
• useraccountnb: user allocation number (if needed)
• resourcemanager: string identifying the manager: 'slurm'|'sge'|'pge'

Parameters marked in bold are required to initialize ClusterSetup object.

Job parameters

• jobname: job name
• queue: job queue where job will be submitted
• nproc: number of processes to allocate (in header)
• logfiledir: directory where all log files will be stored
• nnode: number of nodes to allocate (if needed)
• nthread: number of threads to launch (for each command)

Parameters marked in bold are required to initialize job object.

For example, to allocate 3 nodes and total 24 processes, and running a task with 12 threads could be set with

mpiexec: mpiexec -n {nthread}
nnode: 3
nproc: 24
nthread: 12

Assuming we wish to run task {mpiExec} myprogram -a, on a slurm system this would translate the following entries in the job submission script:

#SBATCH -N 3
#SBATCH -n 24
mpiexec -n 12 myprogram -a

All keywords are read hierarchically from the ClusterSetup, BatchJob and BatchTask objects.

Roadmap

• update selfe example(s)
• clean up old examples