I have, for example, a 4 GPU system. I would like to be able to have 2 users each using MPS on 2 GPUs. In reading the nvidia documentation it looks like I would set CUDA_VISIBLE_DEVICES to specify what GPUs are visible to the nvidia-cuda-mps-control daemon (and then to the server process). Is it then true I would need two instances of the control daemon, each able to see 2 different GPUs in order to have a different user on each pair of GPUs? How would I start 2 instances of nvidia-cuda-mps-control on the same system, can I specify that each use a different port or something?
3 Reset to defaultHow to configure nvidia-cuda-mps-control so that there can be multiple users running on the same system
|
