Hello,
I am a Student currently writing my master thesis on computer vision in the medical field. The institute ordered new computers and I was asked to set them up. I am doing research so far and know what I want roughly. I am only unsure about one thing. My supervisor wants to be able to have multiple users running AI training simultaneously, while always utilizing the most available resources. Additionally there should be no crashes when another training is started. I stubled across multiple methods like vgpu and mig, but I think Time Slicing is the way to go in order to always use maximum resources. I have a few questions about it:
- If I set up 3 users, and have user 1 run a AI Hyperparameter Sweep for lets say 2 weeks. During those two weeks another user wants to also start running other AI trainings. How would this affect the first training? would it just take longer, because its split into time slices? Would it crash?
I appreciate any input on better alternatives or additional knowledge on time slicing. It seems to be pretty straightforward to implement