Dynamic Resource Allocation

Core Dynamic Resource Allocation with structured parameters:

FEATURE STATE: Kubernetes v1.30 [alpha]

Dynamic Resource Allocation with control plane controller:

FEATURE STATE: Kubernetes v1.26 [alpha]

Dynamic resource allocation is an 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. Typically those resources are devices like GPUs.

Third-party resource drivers are responsible for tracking and preparing resources, with allocation of resources handled by Kubernetes via structured parameters (introduced in Kubernetes 1.30). Different kinds of resources support arbitrary parameters for defining requirements and initialization.

When a driver provides a control plane controller, the driver itself handles allocation in cooperation with the Kubernetes scheduler.

Before you begin

Kubernetes v1.31 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 v1.31, check the documentation for that version of Kubernetes.

API

The resource.k8s.io/v1alpha3 API group provides these types:

ResourceClaim
Describes a request for access to resources in the cluster, for use by workloads. For example, if a workload needs an accelerator device with specific properties, this is how that request is expressed. The status stanza tracks whether this claim has been satisfied and what specific resources have been allocated.
ResourceClaimTemplate
Defines the spec and some metadata for creating ResourceClaims. Created by a user when deploying a workload. The per-Pod ResourceClaims are then created and removed by Kubernetes automatically.
DeviceClass
Contains pre-defined selection criteria for certain devices and configuration for them. DeviceClasses are created by a cluster administrator when installing a resource driver. Each request to allocate a device in a ResourceClaim must reference exactly one DeviceClass.
PodSchedulingContext
Used internally by the control plane and resource drivers to coordinate pod scheduling when ResourceClaims need to be allocated for a Pod and those ResourceClaims use a control plane controller.
ResourceSlice
Used with structured parameters to publish information about resources that are available in the cluster.

The developer of a resource driver decides whether they want to handle allocation themselves with a control plane controller or instead rely on allocation through Kubernetes with structured parameters. A custom controller provides more flexibility, but cluster autoscaling is not going to work reliably for node-local resources. Structured parameters enable cluster autoscaling, but might not satisfy all use-cases.

When a driver uses structured parameters, all parameters that select devices are defined in the ResourceClaim and DeviceClass with in-tree types. Configuration parameters can be embedded there as arbitrary JSON objects.

The core/v1 PodSpec defines ResourceClaims that are needed for a Pod in a 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.

apiVersion: resource.k8s.io/v1alpha3
kind: DeviceClass
name: resource.example.com
spec:
  selectors:
  - cel:
      expression: device.driver == "resource-driver.example.com"
---
apiVersion: resource.k8s.io/v1alpha2
kind: ResourceClaimTemplate
metadata:
  name: large-black-cat-claim-template
spec:
  spec:
    devices:
      requests:
      - name: req-0
        deviceClassName: resource.example.com
        selectors:
        - cel:
           expression: |-
              device.attributes["resource-driver.example.com"].color == "black" &&
              device.attributes["resource-driver.example.com"].size == "large"              
–--
apiVersion: v1
kind: Pod
metadata:
  name: pod-with-cats
spec:
  containers:
  - name: container0
    image: ubuntu:20.04
    command: ["sleep", "9999"]
    resources:
      claims:
      - name: cat-0
  - name: container1
    image: ubuntu:20.04
    command: ["sleep", "9999"]
    resources:
      claims:
      - name: cat-1
  resourceClaims:
  - name: cat-0
    resourceClaimTemplateName: large-black-cat-claim-template
  - name: cat-1
    resourceClaimTemplateName: large-black-cat-claim-template

Scheduling

With control plane controller

In contrast to native resources (CPU, RAM) and extended resources (managed by a device plugin, advertised by kubelet), without structured parameters 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.

When a pod gets scheduled, 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.

With structured parameters

When a driver uses structured parameters, the scheduler takes over the responsibility of allocating resources to a ResourceClaim whenever a pod needs them. It does so by retrieving the full list of available resources from ResourceSlice objects, tracking which of those resources have already been allocated to existing ResourceClaims, and then selecting from those resources that remain.

The only kind of supported resources at the moment are devices. A device instance has a name and several attributes and capacities. Devices get selected through CEL expressions which check those attributes and capacities. In addition, the set of selected devices also can be restricted to sets which meet certain constraints.

The chosen resource is recorded in the ResourceClaim status together with any vendor-specific configuration, so when a pod is about to start on a node, the resource driver on the node has all the information it needs to prepare the resource.

By using structured parameters, the scheduler is able to reach a decision without communicating with any DRA resource drivers. It is also able to schedule multiple pods quickly by keeping information about ResourceClaim allocations in memory and writing this information to the ResourceClaim objects in the background while concurrently binding the pod to a node.

Monitoring resources

The kubelet provides a gRPC service to enable discovery of dynamic resources of running Pods. For more information on the gRPC endpoints, see the resource allocation reporting.

Pre-scheduled Pods

When you - or another API client - create a Pod with spec.nodeName already set, the scheduler gets bypassed. If some ResourceClaim needed by that Pod does not exist yet, is not allocated or not reserved for the Pod, then the kubelet will fail to run the Pod and re-check periodically because those requirements might still get fulfilled later.

Such a situation can also arise when support for dynamic resource allocation was not enabled in the scheduler at the time when the Pod got scheduled (version skew, configuration, feature gate, etc.). kube-controller-manager detects this and tries to make the Pod runnable by triggering allocation and/or reserving the required ResourceClaims.

It is better to avoid bypassing the scheduler because a Pod that is assigned to a node blocks normal resources (RAM, CPU) that then cannot be used for other Pods while the Pod is stuck. To make a Pod run on a specific node while still going through the normal scheduling flow, create the Pod with a node selector that exactly matches the desired node:

apiVersion: v1
kind: Pod
metadata:
  name: pod-with-cats
spec:
  nodeSelector:
    kubernetes.io/hostname: name-of-the-intended-node
  ...

You may also be able to mutate the incoming Pod, at admission time, to unset the .spec.nodeName field and to use a node selector instead.

Enabling dynamic resource allocation

Dynamic resource allocation is an alpha feature and only enabled when the DynamicResourceAllocation feature gate and the resource.k8s.io/v1alpha3 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.

When a resource driver uses a control plane controller, then the DRAControlPlaneController feature gate has to be enabled in addition to DynamicResourceAllocation.

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

kubectl get deviceclasses

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

No resources found

If not supported, this error is printed instead:

error: the server doesn't have a resource type "deviceclasses"

A control plane controller is supported when it is possible to create a ResourceClaim where the spec.controller field is set. When the DRAControlPlaneController feature is disabled, that field automatically gets cleared when storing the ResourceClaim.

The default configuration of kube-scheduler enables the "DynamicResources" plugin if and only if the feature gate is enabled and when using the v1 configuration API. 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

Last modified September 04, 2024 at 4:00 PM PST: Fix typo in dynamic-resource-allocation.md (744d7d6f4a)