Understanding dynamic resource allocation in Kubernetes

2026-07-01 · Source: Cloud Native Computing Foundation · Field: Technology & Digital — Cloud Computing & IT Infrastructure, Artificial Intelligence & Machine Learning · Depth: Intermediate, long

Summary

Kubernetes Dynamic Resource Allocation (DRA), now generally available in v1.35, offers a refined approach to managing hardware resources like GPUs. This post, published on July 1, 2026, details its implementation using NVIDIA's maturing dra-driver-nvidia-gpu in a CNTUG Infra Labs environment featuring Kubernetes v1.35.3, Containerd 2.2.2, NVIDIA RTX A5000, and Tesla T10 GPUs. It demonstrates how to install the NVIDIA GPU Operator v26.3.1 and NVIDIA DRA Driver GPU v25.12.0, then explores practical scenarios: sharing a single GPU across containers, prioritizing specific GPU models (e.g., A5000 over T10) in deployments, requesting GPUs based on memory capacity (e.g., >20GiB), and configuring GPU Time Slicing for shared access.

Key takeaway

For AI/ML engineers or DevOps teams managing GPU-intensive workloads on Kubernetes, DRA in v1.35 provides significantly more flexible and precise GPU allocation than the legacy Device Plugin. You can declaratively specify GPU types, memory, or sharing strategies, enabling better resource utilization. Consider migrating existing GPU deployments to DRA for improved management and scaling, but be mindful of how rolling updates interact with ResourceClaimTemplates during transitions.

Key insights

Kubernetes DRA provides granular, declarative GPU allocation, surpassing older Device Plugin limitations.

Principles

• DeviceClass categorizes available hardware.
• ResourceSlice tracks node-specific device pools.
• ResourceClaim/Template manage device requests.

Method

Install NVIDIA GPU Operator and DRA Driver, then define ResourceClaims or ResourceClaimTemplates using `exactly` or `firstAvailable` with CEL expressions for precise device selection.

In practice

• Share a single GPU among multiple containers.
• Prioritize specific GPU models (e.g., A5000).
• Request GPUs based on memory capacity (e.g., >20GiB).

Topics

• Kubernetes
• Dynamic Resource Allocation
• GPU Management
• NVIDIA GPU Operator
• ResourceClaim
• ResourceClaimTemplate

Code references

• kubernetes-sigs/dra-driver-nvidia-gpu

Best for: Machine Learning Engineer, AI Engineer, DevOps Engineer

Related on AIssential

• Get Real-Time Visibility into GPU Usage Across Kubernetes Clusters — The GPU Usage Monitor provides integrated, real-time GPU observability for Kubernetes, preventing over-provisioning and scheduling delays.
• Kubernetes v1.36: Security Defaults Tighten as AI Workload Support Matures — Kubernetes v1.36 enhances security, AI/ML workload management, and API scalability through new stable and beta features.
• Breaking free of a single datacenter: Practical geo-distributed AI operations with the k0smos platforms — Geo-distributed AI operations are viable using Kubernetes-native multi-cluster orchestration for heterogeneous hardware and dynamic resource management.
• Running Large-Scale GPU Workloads on Kubernetes with Slurm — Slinky's slurm-operator integrates Slurm with Kubernetes for scalable, high-performance AI training on GPU infrastructure.
• GPU Clouds, Aggregators, and the New Economics of AI Compute — GPU cloud providers offer diverse services and pricing, necessitating careful evaluation based on security, managed services, and workload portability.
• Maximizing GPU Utilization: Heterogeneous Pipelines with Ray and Kubernetes — Ray optimizes heterogeneous, distributed AI workloads by managing diverse compute resources and handling failures across complex, multi-layered stacks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Cloud Native Computing Foundation.