The NVIDIA Driver Manager is a Kubernetes component that assists in seamless upgrades of NVIDIA drivers on each node of the cluster. This component ensures that all prerequisites are met before driver upgrades can be performed using the NVIDIA GPU Driver. The following actions are performed by this component when an upgrade is required:

Features

• Kernel Module Check: Checks for already installed kernel modules.
• Node Drain: Performs drain on the node, ignoring Daemonset pods.
• Component Eviction: Evicts GPU Operator components like Device-Plugin, GPU Feature Discovery, DCGM Exporter, etc.
• Kernel Module Unload: Unloads kernel modules.
• Filesystem Unmount: Unmounts the driver root filesystem previously mounted on the host under /run/nvidia/driver.
• Node Uncordon: Uncordons the node.

Benefits

• Seamless Upgrades: Allows new driver versions to be easily installed in the Kubernetes cluster.
• Automated Management: Automates the management and upgrade process, reducing manual intervention.
• Enhanced Reliability: Ensures all prerequisites are met for reliable driver upgrades.
• Efficient Resource Management: Improves resource management by handling component eviction and node draining.