Hello all, For years I've been looking for what I might consider to be the holy grail of composable resource allocation in slurm jobs:
* A command that can be run inside of an sbatch script... * ...which immediately and synchronously invokes another sbatch script (which may or may not invoke mpirun in turn)... * ...using a subset of the currently allocated resources. This is the smallest unit of functionality that would compose well with existing tools in UNIX for orchestration. For instance, I could use xargs as a semaphore to let each node work on one input at a time, and for a given input I could have an arbitrarily complex python script decide dynamically what computations to run. Years of Google and manpage searches have continually failed me. * salloc can synchronously run an sbatch script, but as far as I can tell, it cannot make job steps, only jobs. * srun can run sychronously and make job steps, but as far as I can tell, it cannot call a script which calls mpirun (it insists on *replacing* mpirun) * Today I discovered that sbatch can also create job steps (albeit awkwardly, via --jobid), and it can obviously run sbatch scripts... but as far as I can tell, it cannot run synchronously! One can't help but wonder whether this is a deliberate omission or just criminal oversight! Today I snapped and started working on a synchronous wrapper around sbatch; the plan is to use --jobid=$SLURM_JOB_ID, find out the job step id (somehow...), and then sattach to it. I say this knowing it'll probably make your skin crawl. And I ask: What do you think I ought to do instead? Michael
