DRA Consumable Capacity

Without this KEP, device sharing is done by having multiple pods (and/or containers) reference the same resource claim, and that resource claim has allocated the device. This can be considered exclusive or dedicated allocation. With this KEP, independent resource claims (and/or requests within a claim) can allocate shares of resources provided by the same underlying device. This enables resource sharing across pods that are completely unrelated, potentially even across different namespaces. When a device is shared across multiple resource claims, this enhancement enables device resource allocation to be drawn from the device’s overall capacity. It ensures that the total resources consumed by all claim requests remain within the device’s capacity and comply with any defined requestPolicy, such as minimum per-claim resource requirements, if specified. This concept is referred to as consumable capacity. If a request does not specify particular device resource requirements, it implies an expectation of full device capacity. Notably, each of these independent resource claims can still be referenced by one or more pods. However, the device resources allocated to each request are shared without any isolation guarantees among the pods that reference the same request. To achieve this, this KEP introduces - a new device property field to distinguish between devices those can be allocated only once and those can be allocated multiple times, - a capacity-aware scheduling mechanism that allows limiting or guaranteeing the capacity of devices among the resource claims (or requests) those are sharing, - a new capacity requirement field in the device request of the resource claim, - a new consumed capacity field in the allocation result of the resource claim, - a method to associate the allocated device status to the allocation result in the resource claim. With those in place, a resource claim with multiple requests might allocate the same device multiple times. This may or