The NVIDIA MIG Manager is a Kubernetes component capable of repartitioning GPUs into different MIG configurations in an easy and intuitive way. Users simply add a label with their desired MIG configuration to a node, and the MIG manager takes all the steps necessary to make sure it gets applied. This includes shutting down all attached GPU clients, performing the MIG configuration itself, and then bringing those clients back online once the process is complete. Available configurations are stored in a configMap in the cluster, and the MIG manager uses the MIG partition editor to carry out the actual MIG configuration under the hood.

Features

• Simple Label Changes: Manage MIG partitions by adding labels to nodes.
• Automated Configuration: Automatically shuts down GPU clients, performs MIG configuration, and brings clients back online.
• ConfigMap Storage: Stores available configurations in a configMap within the cluster.
• MIG Partition Editor: Utilizes the MIG partition editor for configuration tasks.
• Seamless Integration: Integrates with the NVIDIA GPU Operator for efficient management.

Benefits

• Ease of Use: Simplifies the process of managing MIG partitions.
• Automated Management: Reduces manual intervention with automated configuration steps.
• Enhanced Efficiency: Ensures efficient GPU resource management.
• Reliable Performance: Provides reliable performance with seamless integration.