A GPU Is an Integer to Kubernetes

A GPU enters Kubernetes scheduling through a device plugin. The plugin reports the devices it manages to the kubelet, which advertises them in the node’s status. A plugin might expose a resource named nvidia.com/gpu; a Pod can then request that resource. Kubernetes explains this registration path in its device plugin documentation.

The scheduler counts units

Device plugin resources are extended resources. Kubernetes supports them only as integers and does not overcommit them. A container can request one advertised unit or two, but it cannot request half a unit. The device plugin documentation also says devices cannot be shared between containers through this allocation path.

Suppose a node advertises two healthy units of a GPU resource. A Pod requesting two needs a node that can satisfy both units. Once those units are allocated, scheduling another Pod that requests the same resource requires available units elsewhere. This is a count of the resource the plugin advertises. It is not a reading of how busy the GPU is. The Kubernetes example shows how the advertised count constrains placement.

A GPU limit is also its request

The word limits can be misleading here. For GPUs, Kubernetes says to put the resource in a container’s limits. If requests is omitted, Kubernetes uses the limit as the request. If both are present, their GPU values must match. A GPU request on its own is unsupported. The GPU scheduling guide shows a Pod with a GPU limit of 1.

This matters when a workload appears small enough to share a GPU. A smaller application workload does not produce a fractional device plugin request. The schedulable unit remains the integer resource advertised by the plugin.

What to do

1. Install the GPU driver and corresponding device plugin on the relevant nodes, as the GPU scheduling guide directs.
2. Check the resource name and available count advertised in each node’s status. Use that exact name in the container’s GPU limits.
3. Size the Pod’s integer GPU limit against the available count. If your nodes have different GPU types, use node labels and selectors to place workloads on suitable nodes, following the Kubernetes guidance.