/sci/ - Science & Math

Last month I ran my main research code (written in C++) roughly 900k times to cover a 3D parameter space and generate a set of data for further simulations. This is a normalized distribution of wall-clock time per run completion.

So with a most probably completion time per run of ~4mins, I had to use my university's super cluster. I load balanced these serial runs across many cores to create a task-parallelization to get this shit done in human time.

It took one chunk of 11400 cores for 6hrs, then 3 chunks of 5400 cores for 6 hrs to complete all ~900k runs, totalling roughly 185k CPU-hrs.

Waiting in the cluster's queues while assholes try to put on way too many jobs gets annoying as hell.

Also, I did this entire set computation twice and still have allocation left for another 2 sets of 900k if I want.