The distinguishing characteristic of non-MPI environments is that, while many tasks may be running across an allocation concurrently, they do not communicate via MPI and are not launched by an MPI launcher. Embarrassingly parallel workloads or parameter sweeps are examples of this type of job.

This section describes how concurrent non-MPI tasks are launched and controlled within a Slurm allocation via methods such as srun (from a login node only) and TACC's launcher tool.

 
©   Cornell University  |  Center for Advanced Computing  |  Copyright Statement  |  Inclusivity Statement