Dynamic Resource Allocation

FEATURE STATE: Kubernetes v1.26 [alpha]

Dynamic resource allocation is a new API for requesting and sharing resources between pods and containers inside a pod. It is a generalization of the persistent volumes API for generic resources. Third-party resource drivers are responsible for tracking and allocating resources. Different kinds of resources support arbitrary parameters for defining requirements and initialization.

Before you begin

Kubernetes v1.26 includes cluster-level API support for dynamic resource allocation, but it needs to be enabled explicitly. You also must install a resource driver for specific resources that are meant to be managed using this API. If you are not running Kubernetes v1.26, check the documentation for that version of Kubernetes.

API

The new resource.k8s.io/v1alpha1 API group provides four new types:

ResourceClass

Defines which resource driver handles a certain kind of resource and provides common parameters for it. ResourceClasses are created by a cluster administrator when installing a resource driver.

ResourceClaim

Defines a particular resource instances that is required by a workload. Created by a user (lifecycle managed manually, can be shared between different Pods) or for individual Pods by the control plane based on a ResourceClaimTemplate (automatic lifecycle, typically used by just one Pod).

ResourceClaimTemplate

Defines the spec and some meta data for creating ResourceClaims. Created by a user when deploying a workload.

PodScheduling

Used internally by the control plane and resource drivers to coordinate pod scheduling when ResourceClaims need to be allocated for a Pod.

Parameters for ResourceClass and ResourceClaim are stored in separate objects, typically using the type defined by a CRD that was created when installing a resource driver.

The core/v1 PodSpec defines ResourceClaims that are needed for a Pod in a new resourceClaims field. Entries in that list reference either a ResourceClaim or a ResourceClaimTemplate. When referencing a ResourceClaim, all Pods using this PodSpec (for example, inside a Deployment or StatefulSet) share the same ResourceClaim instance. When referencing a ResourceClaimTemplate, each Pod gets its own instance.

The resources.claims list for container resources defines whether a container gets access to these resource instances, which makes it possible to share resources between one or more containers.

Here is an example for a fictional resource driver. Two ResourceClaim objects will get created for this Pod and each container gets access to one of them.

  1. apiVersion: resource.k8s.io/v1alpha1
  2. kind: ResourceClass
  3. name: resource.example.com
  4. driverName: resource-driver.example.com
  5. ---
  6. apiVersion: cats.resource.example.com/v1
  7. kind: ClaimParameters
  8. name: large-black-cat-claim-parameters
  9. spec:
  10.   color: black
  11.   size: large
  12. ---
  13. apiVersion: resource.k8s.io/v1alpha1
  14. kind: ResourceClaimTemplate
  15. metadata:
  16.   name: large-black-cat-claim-template
  17. spec:
  18.   spec:
  19.     resourceClassName: resource.example.com
  20.     parametersRef:
  21.       apiGroup: cats.resource.example.com
  22.       kind: ClaimParameters
  23.       name: large-black-cat-claim-parameters
  24. –--
  25. apiVersion: v1
  26. kind: Pod
  27. metadata:
  28.   name: pod-with-cats
  29. spec:
  30.   containers:
  31.   - name: container0
  32.     image: ubuntu:20.04
  33.     command: ["sleep", "9999"]
  34.     resources:
  35.       claims:
  36.       - name: cat-0
  37.   - name: container1
  38.     image: ubuntu:20.04
  39.     command: ["sleep", "9999"]
  40.     resources:
  41.       claims:
  42.       - name: cat-1
  43.   resourceClaims:
  44.   - name: cat-0
  45.     source:
  46.       resourceClaimTemplateName: large-black-cat-claim-template
  47.   - name: cat-1
  48.     source:
  49.       resourceClaimTemplateName: large-black-cat-claim-template

Scheduling

In contrast to native resources (CPU, RAM) and extended resources (managed by a device plugin, advertised by kubelet), the scheduler has no knowledge of what dynamic resources are available in a cluster or how they could be split up to satisfy the requirements of a specific ResourceClaim. Resource drivers are responsible for that. They mark ResourceClaims as “allocated” once resources for it are reserved. This also then tells the scheduler where in the cluster a ResourceClaim is available.

ResourceClaims can get allocated as soon as they are created (“immediate allocation”), without considering which Pods will use them. The default is to delay allocation until a Pod gets scheduled which needs the ResourceClaim (i.e. “wait for first consumer”).

In that mode, the scheduler checks all ResourceClaims needed by a Pod and creates a PodScheduling object where it informs the resource drivers responsible for those ResourceClaims about nodes that the scheduler considers suitable for the Pod. The resource drivers respond by excluding nodes that don’t have enough of the driver’s resources left. Once the scheduler has that information, it selects one node and stores that choice in the PodScheduling object. The resource drivers then allocate their ResourceClaims so that the resources will be available on that node. Once that is complete, the Pod gets scheduled.

As part of this process, ResourceClaims also get reserved for the Pod. Currently ResourceClaims can either be used exclusively by a single Pod or an unlimited number of Pods.

One key feature is that Pods do not get scheduled to a node unless all of their resources are allocated and reserved. This avoids the scenario where a Pod gets scheduled onto one node and then cannot run there, which is bad because such a pending Pod also blocks all other resources like RAM or CPU that were set aside for it.

Limitations

The scheduler plugin must be involved in scheduling Pods which use ResourceClaims. Bypassing the scheduler by setting the nodeName field leads to Pods that the kubelet refuses to start because the ResourceClaims are not reserved or not even allocated. It may be possible to remove this limitation in the future.

Enabling dynamic resource allocation

Dynamic resource allocation is an alpha feature and only enabled when the DynamicResourceAllocation feature gate and the resource.k8s.io/v1alpha1 API group are enabled. For details on that, see the --feature-gates and --runtime-config kube-apiserver parameters. kube-scheduler, kube-controller-manager and kubelet also need the feature gate.

A quick check whether a Kubernetes cluster supports the feature is to list ResourceClass objects with:

  1. kubectl get resourceclasses

If your cluster supports dynamic resource allocation, the response is either a list of ResourceClass objects or:

  1. No resources found

If not supported, this error is printed instead:

  1. error: the server doesn't have a resource type "resourceclasses"

The default configuration of kube-scheduler enables the “DynamicResources” plugin if and only if the feature gate is enabled. Custom configurations may have to be modified to include it.

In addition to enabling the feature in the cluster, a resource driver also has to be installed. Please refer to the driver’s documentation for details.

What’s next