We need to do scheduling with GPU resources. We are running a fairly complex processing pipeline where several services use GPU resources for NN inferencing. Depending on the service, anywhere from 1 to about a dozen instances may fit in a single GPU.
We need the scheduler to treat GPU processing and RAM as managed allocatable resources (just like CPU and main RAM). The current state of the Nomad scheduler, together with the nvidia device plugin, seems tantalizingly close to
what we need. It falls down because it treats GPUs as indivisible
resources.