Can we get a set_max_jobs_per_cpu or similar?

[lucasb-eyer]

hi, could we please have a way for ts to consider scheduling multiple jobs on a single GPU automatically?

I understand you didn't like the "free memory" heuristic because of job init time.

May I suggest a "set_max_jobs_per_cpu" kind of flag? If I know that my GPU can fit 3 of my jobs, then I can just set this and things will work, and I'll use my GPUs better.

[lucasb-eyer]

Actually I just realized I can approximate this by, for example, splitting my queue into three queues with three ts running on three different sockets.