Container Image in job spec

[lukasheinrich]

Increasingly folks use Linux containers for encapsulating their workloads and a good standard exists with OCI with HPC compatible runtimes. I haven't found containers being mentioned in the job spec. Is this in scope ?

[hategan]

Hi Lukas,

Any mechanism of starting and running some application through an interface that can be traced to some form of exec*() or wrapper around it is in scope. I would say launching some workload in a container fits the description.

The reason this isn't mentioned explicitly is because it fits not as much in the PSI/J interface, but in the implementation of either a JobExecutor or a JobLauncher. That said, it may be useful to mention, in the introduction of the specification, that the PSI/J abstraction is suitable for launching things in containers should the relevant plugins be written.

[lukasheinrich]

Thanks ! But does the Job know intrinsically about containers or would the user have to eg specify an exec string along the lines of " exec /script.sh" in the job description ?

[hategan]

I am not sure I understand your question. A Job is an abstract object with no knowledge of the exact mechanism used to run it. That mechanism is specified by an implementation of a JobExecutor. When using the API, the programmer would need to use an appropriate instance of (for example) a docker executor which would know how to run the job in a docker container.

[frobnitzem]

Containers should be treated specially by the executor, since the batch queue should be able to do container-specific setup tasks (e.g. distributing the container to the nodes). Note this applies both at job submission time and at srun/launch time.

For slurm, it looks like this would be listed as a job attribute, and as a JobExecutorConfig attribute:
https://slurm.schedmd.com/containers.html#request