DRA: A new era of Kubernetes device management with Dynamic Resource Allocation

This post introduces Dynamic Resource Allocation (DRA), a new Kubernetes-native framework for managing hardware accelerators like GPUs and TPUs at scale.

• •DRA reached stable status in Kubernetes 1.34, replacing the Device Plugin framework which could only express hardware needs as simple integers
• •ResourceSlice API lets drivers publish granular hardware details (memory, cores, architecture, NUMA topology) to the cluster
• •ResourceClaim API allows users to specify precise requirements such as "any GPU with at least 40 GB of VRAM" or inter-device PCIe constraints
• •DRA eliminates manual node pinning by making the Kube-scheduler natively aware of hardware attributes and topology
• •NVIDIA donated its DRA Driver for GPUs and Google donated the DRA driver for TPUs to the Kubernetes community; DRA is now GA in GKE