GPU and ML Workloads on Kubernetes#

Running GPU workloads on Kubernetes requires hardware-aware scheduling that the default scheduler does not provide out of the box. GPUs are expensive – an NVIDIA A100 node costs $3-12/hour on cloud providers – so efficient utilization matters far more than with CPU workloads. This article covers the full stack from device plugin installation through GPU sharing and monitoring.

The NVIDIA Device Plugin#

Kubernetes has no native understanding of GPUs. The NVIDIA device plugin bridges that gap by exposing GPUs as a schedulable resource (nvidia.com/gpu). Without it, the scheduler has no idea which nodes have GPUs or how many are available.

Kubernetes Scheduler: How Pods Get Placed on Nodes#

The scheduler (kube-scheduler) watches for newly created pods that have no node assignment. For each unscheduled pod, the scheduler selects the best node and writes a binding back to the API server. The kubelet on that node then starts the pod. If no node is suitable, the pod stays Pending until conditions change.

The scheduler is the reason pods run where they do. Understanding its internals is essential for diagnosing Pending pods, designing placement constraints, and managing cluster utilization.

Running Windows Workloads on Kubernetes#

Kubernetes supports Windows worker nodes alongside Linux worker nodes in the same cluster. This enables running Windows-native applications – .NET Framework services, IIS-hosted applications, Windows-specific middleware – on Kubernetes without rewriting them for Linux. However, Windows nodes are not interchangeable with Linux nodes. There are fundamental differences in networking, storage, container runtime behavior, and resource management that you must account for.

Core Constraints#

Before adding Windows nodes, understand what is and is not supported:

Taints, Tolerations, and Node Affinity#

Pod scheduling in Kubernetes defaults to “run anywhere there is room.” In production, that is rarely what you want. GPU workloads should land on GPU nodes. System components should not compete with application pods. Nodes being drained should stop accepting new work. Taints, tolerations, and node affinity give you control over where pods run and where they do not.

Taints: Repelling Pods from Nodes#

A taint is applied to a node and tells the scheduler “do not place pods here unless they explicitly tolerate this taint.” Taints have three parts: a key, a value, and an effect.