NVIDIA Run:ai Network Topologies API
The Network Topologies API enables administrators to reflect the hierarchical network topology connectivity of nodes in a data center, such as racks, blocks, and other organizational units, to improve pod scheduling decisions for communication-intensive AI/ML workloads. To support topology-aware scheduling, this API allows administrators to define a multi-level network topology element using ordered Kubernetes node labels that represent the network connectivity of nodes in the data center. Nodes are labeled with the same key-value pairs labels that represent their location in the data center network. These labels are used by the NVIDIA Run:ai Scheduler to prefer scheduling pods on nodes that are "closer" to each other, minimizing communication overhead and optimizing workload performance.