NVIDIA vGPU Device Manager

The NVIDIA vGPU Device Manager manages vGPU devices on a GPU node in a Kubernetes cluster. It defines a schema for declaratively specifying the list of vGPU types one would like to create on the node. The vGPU Device Manager parses this schema and applies the desired config by creating vGPU devices following steps outlined in the NVIDIA vGPU User Guide.

As an example, consider the following configuration for a node with NVIDIA A100 PCIe 40GB cards.

version: v1
    - "A100-40C"

  # NVIDIA A100 PCIe 40GB, C-Series
    - "A100-40C"
    - "A100-20C"
    - "A100-10C"
    - "A100-8C"
    - "A100-5C"
    - "A100-4C"

  # Custom configurations
    - "A100-4C"
    - "A100-5C"
    - "A100-8C"
    - "A100-10C"
    - "A100-20C"
    - "A100-40C"

Each of the sections under vgpu-configs is user-defined, with custom labels used to refer to them. For example, the A100-20C label refers to the vGPU configuration that creates vGPU devices of type A100-20C on all GPUs on the node. Likewise, the A100-4C label refers to the vGPU configuration that creates vGPU devices of type A100-4C on all GPUs on the node.

More than one vGPU type can be associated with a configuration. For example, the A100-small label specifies both the A100-4C and A100-5C vGPU types. If the node has multiple A100 cards, then vGPU devices of both types will be created on the node. More specifically, the vGPU Device Manager will select the vGPU types in a round robin fashion as it creates devices. vGPU devices of type A100-4C get created on the first card, vGPU devices of type A100-5C get created on the second card, vGPU devices of type A100-4C get created on the third card, etc.



Note: Currently this project can only be deployed on Kubernetes, and the only supported way is through the NVIDIA GPU Operator. It is not meant to be run as a standalone component and no CLI utility exists. The instructions below are for deploying the vGPU Device Manager as a standalone DaemonSet, for development purposes.

First, create a vGPU devices configuration file. The example file in examples/ can be used as a starting point:

wget -O config.yaml https://raw.githubusercontent.com/NVIDIA/vgpu-device-manager/main/examples/config-example.yaml

Modify config.yaml as needed. Then, create a ConfigMap for it:

kubectl create configmap vgpu-devices-config --from-file=config.yaml

Deploy the vGPU Device Manager:

kubectl apply -f https://raw.githubusercontent.com/NVIDIA/vgpu-device-manager/main/examples/nvidia-vgpu-device-manager-example.yaml

The example DaemonSet will apply the default vGPU configuration by default. To override and pick a new configuration, label the worker node nvidia.com/vgpu.config=<config>, where <config> is the name of a valid configuration in config.yaml. The vGPU Device Manager continuously watches for changes to this label.


View Github