Generate HBN-BGP deployment files¶
After DPF discovery, generate the management bootstrap and HBN-BGP manifests on the CLI machine:
l8k generate --target dpf --use-case hbn-bgp \
--user-config dpf-site.yaml \
--save-deployment-files ./hbn-bgp
The CLI reads the saved inventory and downloads the selected operator and HBN charts over HTTPS. It does not log in to hosts, DPUs or switches, read credential references, or access Kubernetes. Use the separate DPF deployment workflow to apply the reviewed bundle. The workflow diagram shows every step and failure path.
Do not combine generation with --deploy: generate and review the bundle first,
then run l8k deploy --target dpf with explicit --user-config and the generated
--deployment-files path.
Prepare the inputs¶
The current BF4/HBN-BGP cohort remains beta. Generation success confirms the local bundle and its schemas; it does not qualify hardware or runtime behavior.
Keep the inventory produced by discovery in dpf-site.yaml. It must identify each
selected DPU, its exact host adapter, verified BMC administrator access and PSID,
plus the management cluster identity and Kubernetes version. If the selection has
changed, rerun discovery.
Complete these sections in the saved site file:
| Input | Responsibility |
|---|---|
provisioning.bootstrap.tenant.vip and .interface |
Reserve the tenant API address and identify the management interface on its control-plane nodes. |
provisioning.bootstrap.managementAPIIP and .managementAPIPort |
Needed when the saved API endpoint is not a literal, reachable IP address. |
provisioning.software.osISO |
Supply a credential-free HTTPS URL and SHA-256 digest for the selected DPU OS ISO. |
provisioning.software.pldm.<PSID> |
Supply a matching firmware bundle URL and SHA-256 digest for every selected PSID. |
provisioning.flavor |
Supply the concrete DPUFlavor spec for this hardware and HBN datapath. |
provisioning.resources |
Optional overrides of HBN container resources. Defaults are memory: 6Gi and nvidia.com/bf_sf: "4". |
The flavor controls device boot, NIC and datapath settings during provisioning. Use settings appropriate to the selected hardware and software cohort. Schema validation alone cannot establish their runtime compatibility.
Resource quantities form a flat map; this HBN chart uses the same quantities for requests and limits. For example, to raise the memory allocation while retaining the default SF allocation:
The profile sets provisioning.bootstrap.bmcFactoryResetPolicy: Never, which
preserves existing BMC network settings. The current native HBN deployment path
requires Never. Although the general configuration schema and generation can
accept and render OnInitialization, native deployment rejects that policy
before enrollment. Do not rely on DHCP recovery for this path.
Review bootstrap and HBN defaults. Tenant pod and service ranges belong to the independent DPU cluster. The selected release fixes the tenant service range; choose a nonoverlapping pod range in your site configuration. The default ASN allocation fits five DPUs; a larger estate requires an appropriate base/stride override.
The native workflow checks every directed host pair sequentially, with three ordinary
and three DF packets at one-second intervals. Packet spacing alone requires 4 × N × (N
− 1) seconds for N selected physical hosts. Discovery and generation reject a matrix
whose lower bound cannot fit the two-hour native maximum. This rejects 43 or more
selected hosts under the current protocol; fewer hosts are not guaranteed to finish
because provisioning, SSH and API checks take additional time. Generation checks this
deterministic native-plan limit without measuring live readiness.
Plan host workload addresses¶
Each DPU receives a subnet from the workload address pool. The default pool is
10.0.120.0/22, divided into /29 subnets. HBN uses address index 1 in its
allocated subnet as the gateway, and Launch Kit reserves index 2 for the host.
For example, an allocation of 10.0.120.8/29 gives HBN gateway 10.0.120.9 and
reserves 10.0.120.10 for its host.
Override these indexes together when necessary:
The indexes must be different usable addresses within each subnet. Generation excludes the host address from ordinary dynamic IPAM allocation in every possible subnet; it does not predict which DPU will receive which subnet. The pool may contain between 2 and 4096 such subnets. The per-DPU prefix must be more specific than the pool prefix, so the workload-pool route differs from the host's connected subnet. A larger pool needs a broader per-DPU subnet or a smaller workload range.
The host address can only be configured after DPF has allocated that DPU's subnet and its host-facing interface has been identified. The intended host route covers the workload pool through that DPU's HBN gateway. It preserves the existing host default route and management network. Generation prepares the address reservations; the deploy command configures the host only after observing its actual DPF allocation and verifying the selected interface.
The profile also carries MTUs through the complete datapath. The DPF high-speed network and physical service-chain ports use the uplink MTU; the host-facing service-chain port and host workload interface use the workload MTU. These values do not change the independent control-plane MTU.
Review the output¶
The output directory contains:
- Management namespaces, the DPF operator chart and its rendered preview.
- DPFOperatorConfig, the tenant DPUCluster and one exact BMC discovery target per selected DPU.
- Software/flavor definitions, address pools, service interfaces, HBN service configuration and DPUDeployment.
- The verified chart archives, original source YAML and
.l8k/resolved-config.yaml. - An index recording the ordered stages and required deployment checks.
The physical and host-facing interface definitions carry selectors derived from the intended DPU cluster and selected DPUs' node names. Generation does not check whether live labels uniquely identify those targets. HBN's resource requests are recorded both for its container and for DPF's check that the services fit the chosen DPU flavor. Generation verifies the manifest shapes; actual available resources and interface readiness still require deployment checks. Regenerate older preview bundles that omit these selectors or DPF resource requirements; editing their manifests invalidates the retained deployment plan.
The resolved configuration records the effective defaults and overrides, including the operator release name and namespace, each selected DPU's BMC address, the BMC reset policy, the release-selected tenant Kubernetes version, and the selected OS and firmware URLs and checksums. The management cluster and DPU cluster have separate Kubernetes versions; generation does not upgrade the management cluster. Keep it with the manifests so later steps can verify the selected devices and software without resolving the inputs again. To change the plan, edit the site configuration and generate a new directory; do not edit files inside an existing bundle.
Generation validates the full bundle before publishing the directory. Repeating an identical generation leaves that directory unchanged. Use a new destination for changed inputs; generation does not replace a different or incomplete bundle.
A successful result means the artifacts are ready for review. Discovery blockers remain unresolved until the environment is corrected and checked again. BGP adjacency, host traffic, tenant health and consumed artifact versions need live validation after deployment; they cannot be established by rendering YAML.