DRA Partitionable Devices

One of the original use-cases for Dynamic Resource Allocation (DRA) was the ability to dynamically allocate partitions of a full device (in addition to the full device itself). Whereas the traditional device plugin API forces users to statically partition devices ahead of time, DRA was designed to allow those partitions to be created on demand. This leads to increased resource utilization as the size of each partitioned device can be matched in real-time to the workload requesting it. Devices represented in DRA don't necessarily have to be a single unit connected to a single machine, but can also be a logical device comprised of multiple devices connected to multiple machines. Similar to the single device partitioning, users might require either the full multi-host device or a subset. As DRA has evolved from what we now call "classic" DRA to "structured parameters" this ability to dynamically partition devices has been lost. This KEP proposes a method to bring this capability back within the framework that "structured parameters" provides. Additionally, it provides primitives to represent both full devices and their partitions in a more compact way than is possible today. Note that the extensions proposed in this KEP are completely transparent to the end-user. In 1.31, it is already possible to select a device partition via a ResourceClaim, so long as the device vendor has pre-partitioned its devices and advertised them as such inside a ResourceSlice. The user-facing mechanism to select these devices doesn't change with the extensions proposed in this KEP. Instead, the proposed extensions provide the ability for a vendor to advertise "overlapping" partitions, such that the scheduler will never allocate conflicting ones at the same time. This, in turn, gives the vendor the opportunity to dynamically create these partitions after they have been allocated, rather than requiring them to be created before.