Join GitHub today
Great! This is awesome, I'm so glad you guys jumped on this grenade.
Sorry to rehash what we already talked about, but I am a slow thinker: I was wondering if you thought it might be possible to be able to be able to still use the num_jobs and cores_per_job fields rather than specifying the number of machines. For example I am thinking of a situation like this: you have a cluster of two nodes with 64 cores each and you want to run an IPython cluster with 4 engines and each engine be dedicated 12 cores because you are calling out to subprocess with a command that runs as a threaded application, and you want to give each call out 12 cores to work with. -N tells it to dedicate it on 2 separate nodes, which isn't exactly what we want since now we need to know about the machines in our cluster. Ideally, in our 2x64 core cluster above, if one node is empty it should run all 4 engines, each with 12 cores dedicated to it, on one node and fill it up. If each node is half full, it should run 2 engines on one node and 2 on the other. If one node is completely full and the other is half full it should stay pending.
Is there a way we can have SLURM handle that and figure out how many machines to put the job on? Ideally we shouldn't care about anything other than for each engine, it has 12 cores dedicated on the same node, that is all and not how many different nodes they are on. Does srun -n 4 --cpus-per-task=12 do that?