Skip to content

Concepts

Launch Kit translates discovered server hardware and deployment intent into Network Operator resources. These terms describe the data model used across discovery, generation, deployment, and validation.

Hardware

Term Meaning
Physical function (PF) A physical PCI function exposed by a NIC. Launch Kit records its PCI address, device ID, network interface, RDMA device, NUMA node, traffic role, and rail.
Virtual function (VF) A virtual PCI function created from an SR-IOV PF and assigned to a workload.
East-west The workload data-plane path between cluster nodes. East-west PFs determine rails and generated secondary networks.
North-south The path from a node toward management, storage, or external networks. These PFs remain in discovery data but do not create workload rails.
Rail One independent east-west network path. Multirail profiles render resources per rail.
Plane A Spectrum-X subdivision across rails. Plane membership comes from Spectrum-X topology data rather than standard-profile rail discovery.
Fabric The configured link layer, either Ethernet or InfiniBand. Fabric does not describe the Kubernetes device-exposure model.

Groups And Buckets

A source group is one clusterConfig entry produced by discovery. It represents nodes with the same resolved machine type, GPU type, and NIC layout. Its selector uses the nvidia.kubernetes-launch-kit.machine label.

A render bucket is a generation-time grouping of compatible source groups. Groups with the same GPU type and east-west rail count can share bucket-scoped resources. The merged selector uses the nvidia.kubernetes-launch-kit.gpu label.

The distinction matters in heterogeneous clusters: interface naming can remain source-specific while networks, pools, and workloads are shared by a render bucket.

Deployment Types

Deployment type Workload access
sriov Creates VFs and exposes dedicated devices through SR-IOV networks and device-plugin resources.
host_device Moves an entire physical interface into a workload network namespace.
rdma_shared Shares the host RDMA device while a Macvlan or IPoIB interface provides workload connectivity.

Fabric and deployment type are independent inputs. For example, sriov can use Ethernet or InfiniBand, while rdma_shared maps to Macvlan on Ethernet and IPoIB on InfiniBand.

Profiles And Presets

A deployment profile selects the manifest family from fabric, deployment type, multirail, routing, and optional Spectrum-X settings.

A topology preset is a certified hardware description for a known (machineType, gpuType) pair. A matching preset can replace discovery heuristics for traffic, rail, NUMA, and GPU affinity only when the discovered PF topology matches completely. A deviation remains live-discovered state and blocks validation acceptance.

Workflow State

Artifact Owner Role
cluster-config.yaml Discovery and user Hardware inventory plus resolved deployment intent.
deployment/network-operator/values.yaml Generation Helm values for the Network Operator release.
Generated YAML manifests Generation Declarative resources applied in dependency order.
Validation HTML report Validation Acceptance evidence for release, live resource state, topology, and data-plane checks.

For the detailed behavior behind these concepts, continue with Cluster Discovery, Manifest Generation, and Heterogeneous Clusters.