How Do I Configure a Pod to Use the Acceleration Capability of a GPU Node?

Problem Description

I have purchased a GPU node, but the operating speed is still slow. How do I configure the pod to use the acceleration capability of the GPU node?

Solution

Solution 1:

You are advised to remove the unschedulable taints from the GPU nodes in the cluster, so that the GPU plug-in driver can be properly installed. In addition, you need to install the GPU driver of a later version.

If a container is not deployed on a GPU node in your cluster, you can configure affinity and anti-affinity policies to prevent the container from being scheduled to the GPU node.

Solution 2:

You are advised to install the GPU driver of a later version and use kubectl to update the GPU plug-in configuration. Add the following configuration:

tolerations:
- operator: "Exists"

After the configuration is added, the GPU plug-in driver can be properly installed on the GPU node with a taint.