NVIDIA
Can we support set permission (read/write/mknod) for GPU
In our case. some container will write data to GPU to finish ML job. But some container only monitor the GPU resources. We want to only give the read permission for monitor container. Can we specification the device permission through nvidia-docker?
For docker I saw we can set the device permission through --device=<Host device>:<Container device>:<Permission mode>
The only device file that is actually written to is /dev/nvidiactl and you need to be able to perform both read and writes from this even in the monitoring case. No data is ever written to the respective /dev/nvidia0,1,2,etc device nodes, but write permission needs to be set on them for the kernel to allow access to them through /dev/nvidiactl (even if just monitoring them).
