Make submitter an independent pip package

[loeiten]

The submitter package is an independent unit of code, and can be refactored out.
Suggestion for name cluster-submitter

[loeiten]

Has this been done before?

Partly:

• There does not seem to be a "the package" for cluster submission of costume jobs through python
• Most packages looks to be done half-hearted with few collaborators or maintainers, and almost no documentation
• Some packages seems to have solved the problem of submitting jobs to several different clusters, but the submission routines are bundled in their main program

PBS

• pbs - Few contributors. Possibility to reuse the pbs.job module, but not documented other than README.md
• pyqueue - Not documented
• pbs-python - Not documented - one user

SLURM

• Simple Slurm - Only 4 commits, but looks like it has good documentation, and that it is easy to use. May be reused
• submitit - some documentation and examples provided. Small community. May be reused
• SlurmQueen - For running machine learning experiments with through SLURM
• slurm-toys - Scraped off stuff from other GitHub places
• slurmpy - Very minimal wrapper around the submission job
• slurm-pipeline - Submits jobs through json with possibility of command-line configuration
• pyssub - One user. Submit jobs through json files
• PySlurm - Some collaborators. Only API doc available
• slurmjobs - Looks like most suited for machine learning experiments
• admiral - small code base, no documentation

Both

• Dask - well supported code, looks like made for analytics on large datasets. The submitter functions are found in job queue, but doesn't seem to support submitting a job (rather interactive use)
• jobsubmitter - aimed to connect from Jupyter notebook and submit from there. Lacks documentation
• SoS - Script of scripts - huge system somewhat messy code quality. Maybe parts of sos/utils and (sos-pbs)[https://github.com/vatlab/sos-pbs/tree/master/src/sos_pbs] can be reusable. The usage seems to be well documented
• radical-saga - Aiming for submission through ssh. Seems to be ok supported, and is somewhat popular. Good documentation. Might be possible to reuse
• fyrd - No longer actively supported. Small number of users. Looks like it's aiming for doing many tasks at once through interactive python use

Related

• Resource management - Ideas on how to implement submission through python
• hpc05 - connect to PBS or SLURM cluster with headnode over ssh
• pysge - run on SGE batch cluster with tqdm progress bars
• Remote SGE - One user. Submit to SGE