intel-resource-drivers-for-kubernetes

Intel GPU resource driver

CAUTION: This is an beta / non-production software, do not use on production clusters.

Glossary

• DRA https://github.com/kubernetes/enhancements/tree/master/keps/sig-node/3063-dynamic-resource-allocation
• CDI https://github.com/container-orchestrated-devices/container-device-interface/
• K8s https://github.com/kubernetes/kubernetes.git

About resource driver

Intel GPU resource driver is a better alternative for Intel GPU device plugin, facilitating workload offloading by providing access to GPU devices available on worker nodes of Kubernetes cluster.

Supported GPU devices:

• Intel® Data Center GPU Flex Series
• Intel® Data Center GPU Max Series
• any integrated Intel GPU supported by the host kernel

About Dynamic Resource Allocation

Dynamic Resource Allocation (DRA) is a resource management framework in Kubernetes (1.26+), that allows management of special resources in cluster (typically HW accelerators) by vendor-provided resource drivers (typically a controller and a node-agent / kubelet-plugin) in a common way.

Resource drivers are meant to handle discovery, allocation, accounting of specific resources as well as their preparation for Pod before Pod startup and cleanup after the Pod has completed successfully and the resource is no longer needed. More info is in the KEP

Intel GPU resource driver consists of the controller and kubelet plugin. Controller makes allocation decisions and kubelet plugin ensures that the allocated GPUs and SR-IOV Virtual Functions are prepared and available for Pods

Requirements

• Kubernetes 1.26, with DynamicResourceAllocation feature-flag enabled, and other cluster parameters Container runtime needs to support CDI:
• CRI-O at least v1.23.0
• Containerd at least v1.7 (any release candidate will do)

Supported Kubernetes Versions

Supported Kubernetes versions are listed below:

How to deploy and use

Building resource-driver

Build a container

make container-build

NB: It is recommended to have a local registry to deploy from, so all cluster nodes can access same image and no need to build same image on evey node.