Generated at build time from charts/racora-cu/README.md (helm-docs: values.yaml + README.md.gotmpl); edit the source, not this page.

racora-cu

The three CU planes of the Racora gNB packaged together. All three deploy into the centralized-unit namespace.

What this installs

Plane	Resources
CU-CP	cu-cp Deployment (with optional RANN-P sidecar) · cu-cp Service (F1AP, E1AP, metrics-ws). Its runtime config is the controller-generated nrcell-cu-cp-config ConfigMap (AMF address, max_nof_dus, mobility, admin intent) on top of the vendored base ocudu-base-configs
CU-UP	cu-up-config ConfigMap · cu-up Deployment
CU-IP	cuip ServiceAccount + ClusterRole + ClusterRoleBinding (read NRCells) · cuip-env ConfigMap · cuip Deployment (GPU, Flight on port 8815) · cuip Service

CU-CP and CU-UP are independently togglable (cuCp.enabled, cuUp.enabled — both default true); CU-IP is integral to the distribution and always renders (no enable flag). The AMF the CU-CP dials is the umbrella's global.core.amfAddr (an external core), rendered into the overlay by the controller — this chart carries no AMF value.

All three planes ship dormant (replicaCount: 0): the controller scales the CU plane to 1 when the first NRCell is declared and dissolves it back to 0 when the last cell is deleted.

Requires

• racora-crds — NRCell CRD (CU-IP reads this)
• racora-base — centralized-unit namespace
• racora-controller — generates the nrcell-cu-cp-config ConfigMap that CU-CP's init container waits on (without the controller, CU-CP boots are blocked indefinitely on the wait-for-cu-cp-config init)

Install

helm install racora-cu ./charts/racora-cu

Values

Key	Type	Default	Description
namespace	string	"centralized-unit"	Namespace the three CU planes are deployed into. Must already exist (racora-base creates it).
imageRegistry	string	""	Standard image registry for the gNB and CU-IP images. Empty = registry-agnostic (bare image names per workspace convention). Override per-environment to point at the registry you publish to.
cuCp.enabled	bool	true	Render the CU-CP plane.
cuCp.image.repository	string	"gnb"	gNB image name.
cuCp.image.tag	string	"v0.4.6"	gNB image tag; pinned from gnb in versions.yaml at bake time (the tree carries a fallback). One image runs CU-CP, CU-UP and every DU.
cuCp.image.pullPolicy	string	"IfNotPresent"	Image pull policy — IfNotPresent, because the images ship with the distribution.
cuCp.replicaCount	int	0	Replicas at install: 0, dormant. The controller owns this field and scales the plane 0↔1 with cell existence — do not set it.
cuCp.substrateAddr	string	"cuip.centralized-unit.svc.cluster.local:8815"	Service DNS for the CU-IP substrate (Flight server). Used by the RANN-P forwarder sidecar AND CU-CP itself when RANN-P is enabled.
cuCp.initWaitForAmfSeconds	int	15	Init delay before cu-cp tries to reach the AMF. SCTP can't be probed by a port probe, so this is a fixed sleep matching the open5gs boot time.
cuCp.baseConfigHostPath	string	""	OCUDU base config source. By default the chart ships its own ocudu-base-configs ConfigMap with cu_cp.yml + cu_up.yml vendored from upstream OCUDU configs/ (BSD-3-Clause-Open-MPI) and mounts it at /etc/ocudu. Zero host-side prep needed for a default install. Set this to a non-empty path to mount /etc/ocudu from a host directory instead — escape hatch for operators who want to fully replace the chart-provided defaults.
cuCp.rannP.enabled	bool	true	Enable the RANN-P sidecar that ships measurements from CU-CP's file sink to the CU-IP substrate via Arrow Flight do_put.
cuCp.rannP.image.repository	string	"rann-p-forwarder"	RANN-P forwarder sidecar image name.
cuCp.rannP.image.tag	string	"v0.1.1"	RANN-P forwarder sidecar image tag; pinned from rann_p_forwarder in versions.yaml.
cuCp.rannP.image.pullPolicy	string	"IfNotPresent"	Image pull policy — IfNotPresent, because the images ship with the distribution.
cuCp.rannP.batchSize	int	50	RANN-P batch size, shared by the CU-CP's own sink and the forwarder sidecar (both read it from this one value).
cuCp.resources	object	{}	Pod resources for the CU-CP container.
cuUp.enabled	bool	true	Render the CU-UP plane.
cuUp.image.repository	string	"gnb"	gNB image name.
cuUp.image.tag	string	"v0.4.6"	gNB image tag; pinned from gnb in versions.yaml at bake time; keep equal to cuCp.image.tag.
cuUp.image.pullPolicy	string	"IfNotPresent"	Image pull policy — IfNotPresent, because the images ship with the distribution.
cuUp.replicaCount	int	0	Replicas at install: 0, dormant. The controller owns this field and scales the plane 0↔1 with cell existence — do not set it.
cuUp.cuCpAddrs	string	"cu-cp.centralized-unit.svc.cluster.local"	CU-CP service DNS — CU-UP connects to this on E1AP.
cuUp.initWaitForCuCpSeconds	int	20	Init delay before cu-up tries to reach cu-cp (which itself waits for AMF).
cuUp.baseConfigHostPath	string	""	Same as cuCp.baseConfigHostPath — defaults to empty (chart-shipped ConfigMap from files/cu_up.yml). Set a host path to override.
cuUp.resources	object	{}	Pod resources for the CU-UP container.
cuIp.runtimeClassName	string	"nvidia"	Container runtime that injects the NVIDIA driver. k3s registers the "nvidia" RuntimeClass automatically when the NVIDIA container toolkit is present on the node. Override for distributions that name it differently.
cuIp.image.repository	string	"cuip"	CU-IP image name.
cuIp.image.tag	string	"v0.3.0"	CU-IP image tag; pinned from cuip in versions.yaml at bake time.
cuIp.image.pullPolicy	string	"IfNotPresent"	Image pull policy — IfNotPresent, because the images ship with the distribution.
cuIp.replicaCount	int	0	Replicas at install: 0, dormant. The controller owns this field and scales the plane 0↔1 with cell existence — do not set it.
cuIp.crdNamespace	string	"racora-system"	Where the substrate watches for NRCell CRs. Matches the controller's crdNamespace.
cuIp.gpuCount	int	1	CU-IP runs on GPU. nvidia.com/gpu resource. Set to 0 to disable GPU (useful for batch-mode standalone substrate testing).
cuIp.mode	string	"streaming"	Streaming mode is the primary path (V2: live RANN-P + CRD watcher). Set to "batch" + provide UE_REPORTS_PATH / CELL_POSITIONING_PATH via extraEnv to run offline against Parquet files.
cuIp.flightPort	int	8815	Arrow Flight port the substrate serves on (FLIGHT_PORT); the cuip Service exposes it and cuCp.substrateAddr must match.
cuIp.graphAlpha	string	"0.1"	Graph edge-weight smoothing factor (GRAPH_ALPHA), as a string.
cuIp.streamGraphMode	string	"time_ema"	Streaming graph update mode (STREAM_GRAPH_MODE); time_ema decays edges with the half-life below.
cuIp.streamHalfLifeS	string	"60.0"	Edge-weight half-life in seconds for time_ema (STREAM_HALF_LIFE_S), as a string.
cuIp.reportsWindowSize	string	"50000"	Number of measurement reports kept in the streaming window (REPORTS_WINDOW_SIZE), as a string.
cuIp.engines	string	"pci,anr"	Intelligence functions the substrate runs (CUIP_ENGINES, comma- separated; unknown names fail the pod fast). Matches the image default — listed here so an operator can disable a function (e.g. "pci" alone) without image knowledge.
cuIp.engineConfig	object	{}	Per-function engine config, keyed by function name; rendered as the CUIP_ENGINE_CONFIG JSON env when non-empty. Example: engineConfig: anr: theta_add: 0.4 ownership: report-only
cuIp.dataHostPath	string	""	Substrate /data path source. Empty by default — the public substrate doesn't ship with any baked /data content. Set this to a non-empty host path to mount your own /data directory via hostPath when your intelligence functions need data files.
cuIp.serviceAccount.create	bool	true	Create the cuip ServiceAccount, ClusterRole and binding (read NRCells).
cuIp.serviceAccount.name	string	""	ServiceAccount name; empty = the chart-generated default.
cuIp.extraEnv	list	[]	Extra env vars for substrate — useful for batch mode or for overriding any of the substrate-tunable knobs not surfaced above.
cuIp.resources	object	{}	Pod resources for the CU-IP container (the GPU request comes from gpuCount).
controlPlanePin	bool	true	Pod scheduling (applied to every plane unless overridden per-plane). The CU plane pins to the control node: Racora's model is "first node is the brain" (CU + core), additional nodes are DU/RF capacity — so the CU plane must not drift onto an RT-tuned DU worker. controlPlanePin adds node-role.kubernetes.io/control-plane=true to every plane's nodeSelector; the k3s platform labels its server that way (schedulable, untainted), so on a single-node cluster this is a no-op. Platforms whose control planes are tainted NoSchedule (kubeadm/CAPI, OpenShift) set it false and pin with nodeSelector instead (platforms/kubernetes/values.yaml). cuip additionally merges racora.io/gpu-ready when gpuCount>0.
nodeSelector	object	{}	Node selector for the CU planes (every plane unless overridden per plane).
tolerations	list	[]	Tolerations for the CU planes (every plane unless overridden per plane).
affinity	object	{}	Affinity rules for the CU planes (every plane unless overridden per plane).

hostPath caveat

CU-CP and CU-UP mount the OCUDU base configs from cuCp.baseConfigHostPath (empty by default; the chart ships an ocudu-base-configs ConfigMap with upstream OCUDU defaults). Set this to a host path to override.

cuIp.dataHostPath is empty by default. Set it to a host path when your intelligence functions need data files mounted into the substrate's /data directory.

Verify

kubectl get pods -n centralized-unit
# Expected once the controller has reconciled at least one NRCell:
#   cu-cp-xxx     Running
#   cu-up-xxx     Running
#   cuip-xxx      Running

kubectl exec -n centralized-unit deploy/cuip -- ss -tlnp | grep 8815