Why use Capacity instead of Allocatable?

[fighterhit]

Why use node.Status.Capacity instead of node.Status.Allocatable? Doesn't the number of Allocatable determine whether it is schedulable? In this case, I feel that there will be a problem when node.Status.Allocatable is less than node.Status.Capacity((perhaps due to health check by device plugin),), because at this time gpu-admission thinks that there is still a GPU that can be scheduled, but in fact there is no GPU. Maybe there will be problems when the container is allocated GPU on the node?

gpu-admission/pkg/util/util.go

Lines 100 to 117 in 47d56ae

	func GetCapacityOfNode(node *v1.Node, resourceName string) int {
	val, ok := node.Status.Capacity[v1.ResourceName(resourceName)]
	if !ok {
	return 0
	}
	return int(val.Value())
	}
	// GetGPUDeviceCountOfNode returns the number of GPU devices
	func GetGPUDeviceCountOfNode(node *v1.Node) int {
	val, ok := node.Status.Capacity[VCoreAnnotation]
	if !ok {
	return 0
	}
	return int(val.Value()) / HundredCore
	}

[cc1416619381]

I think it uses capacity as total resources, and it will update the used resources with AddUsedResources interface while the correspondings resources are allocated in AllocateOne interface. Thus, we can maintain the infomation about the remaining resources on each node and each device.

If uses node.Status.Allocatable , we can not know about how many resources each device has on each node.