Feature Request: MPS memory control / oversubscription

[adammw]

Current device plugin MPS daemonset sets set_default_device_pinned_mem_limit to totalMemory/replicas/1024/1024. This feature requests to allow "burst" capacity where we would like to oversubscribe the memory limit larger than the fair division of total memory on the assumption that not all processes will use to their limit.

Possible configuration could be via the mps configmap setting a ratio of oversubscription to multiply the total memory by.