facebookincubator
Support Slurm Heterogeneous Job
Does submitit support Slurm Heterogeneous Job? If so, how can we submit heterogeneous job? If not, could you enhance the code to support it?
it's not supported atm, and from an API perspective I'm not sure how to handle this. currently the api assumes there is one configuration per job, while here you want several configurations in the same job. Not impossible, but also non trivial. What's the use case ? can you approximate this by starting two jobs ?
For example, if you want to start a distributed job running on two different GPUs or CPUs with different specs, and they need to communicate with each other and they cannot be started separately.
Here is one example: https://research-computing.git-pages.rit.edu/docs/slurm_tutorial_2.html https://slurm.schedmd.com/heterogeneous_jobs.html#submitting