Kubernetes GPU Scheduling — DRA, Gang Scheduling, and Getting Pods on the Right GPUs

Dynamic Resource Allocation (DRA) is the modern way to request accelerators flexibly in Kubernetes. Its Core APIs have been GA since v1.34, and extended-resource support reached GA in v1.37, so DRA drivers can satisfy accelerator-style requests without a separate device plugin in many cases. For complex GPU/accelerator needs it is increasingly the right model; we help you decide based on your driver support and workload and avoid mixing models in ways that cause Pending Pods.

As of Kubernetes v1.37 (released August 2026), the Workload/PodGroup APIs and gang scheduling graduated to Beta, along with Workload-Aware Preemption and shared DRA ResourceClaims for PodGroups — one step from GA. Importantly, gang scheduling ships disabled by default in v1.37, so it must be explicitly enabled. We help you turn it on safely or use a batch scheduler (Kueue, Volcano) as appropriate, so multi-node training and inference schedule all-or-nothing instead of deadlocking.

CompositePodGroup is a new API introduced in v1.37 for expressing hierarchical, multi-level topology constraints, gang scheduling, and preemption policies for complex, heterogeneous groups of Pods — for example multi-role distributed workloads that need different placement rules per role. We help you model real distributed AI jobs with it where it fits.

Common causes: no allocatable GPU on matching nodes, taints without tolerations, a DRA ResourceClaim that cannot be satisfied, node-affinity/selector mismatches, insufficient quota, or a gang/PodGroup waiting for all members. We diagnose from scheduler events and node capacity and fix the real cause — see our GPU Pod Pending troubleshooting guide.

Message us on WhatsApp with your workloads (training/inference, single vs multi-node), GPU types, and Kubernetes version. We will design scheduling that keeps GPUs busy and jobs reliable — same-day.