IP / MPLS / Transport Domain
Last updated 13 Oct 2026
At a glance
Every other domain eventually runs over IP/MPLS and transport. It is the most heavily orchestrated domain in a telco, and the only one with standard YANG service models in production use. The domain controller exposes services such as L3VPN upward, hides per-vendor device detail below, and selects paths from live topology.
The picture
One L3VPN order, from customer-facing service model to device configuration, with topology flowing back up
1 · Service order
CFS:L3VPN
- sites
- 2
- access
- 1 Gbps
- transport class
- gold
No path, no VRF name, no vendor. The catalog stays abstract.
2 · Service model
What the customer sees
Sites, attachment circuits, SLA
IETF L3SM (RFC 8299)
What the network must do
VRFs, route distinguishers, route targets, BGP-VPN sessions
IETF L3NM (RFC 9182)
3 · Domain controller
Controller
- Renders L3NM into per-vendor device config
- Path computation picks an SR path for class gold
- Commits across both PEs as one transaction
↑ BGP-LS: live topology, TE attributes, SR labels
↓ PCEP: computed path to the headend PE
4 · Network
PE · site A
VRF, BGP-VPN, QoS
PE · site B
VRF, BGP-VPN, QoS
P routers
no per-customer state
Route reflector
control plane only
Only the PEs change for this order. The core is built capacity.
How it works
Elements in the domain
| Element | Role | Typical integration |
|---|---|---|
| PE (provider edge) router | Customer-facing termination: VRFs, BGP peering, QoS, SR endpoints | NETCONF/YANG, gNMI; CLI where models fall short |
| P (provider) router | Core forwarding; no per-customer state | NETCONF/YANG; BGP-LS for topology |
| Route reflector | BGP control-plane scale; no data-plane role | Mostly read-only |
| Transponders / ROADM | Optical layer: wavelengths, fibre routing | NETCONF/YANG (OpenROADM-aligned) or proprietary EMS |
| Path computation element | Central SR / RSVP-TE path computation with state | PCEP, NETCONF |
| Aggregation / metro routers | Subscriber aggregation, BNG handoff, metro Ethernet | NETCONF/YANG |
IP/MPLS, transport and optical are treated as one orchestration domain because their service models compose: an L3VPN depends on a transport class, which depends on optical wavelength availability. Splitting them across separate orchestrators forces the cross-domain layer above to reason about layer 1, 2 and 3 interactions the domain controller should hide.
What the domain exposes upward
| Abstraction | Standard model | Hides |
|---|---|---|
| L3VPN | IETF L3SM (RFC 8299), L3NM (RFC 9182) | VRF config, BGP-VPN sessions, route-target policy |
| L2VPN / EVPN (E-Line, E-LAN, E-Tree) | IETF L2NM (RFC 9291) | EVPN BGP config, MAC-VRF, MTU, broadcast handling |
| IP transit / Internet | OpenConfig network-instance and BGP | Peer config, route policy, RPKI, prefix lists |
| Optical wavelength / OTN | IETF L1CSM, OpenROADM | Port maps, optical reach, amplifier settings |
| Segment-routed transport | IETF SR policy, OpenConfig SR | SID stacks, PCE state |
| Transport slice | IETF network slice models (still maturing) | Slice realisation across SR, RSVP-TE, optical |
A service-level northbound interface lets the cross-domain orchestrator compose services without knowing which vendor’s PE is at the other end.
Service model and network model
L3SM describes the service the customer sees: sites, attachment circuits, SLA. L3NM describes what the network does to deliver it: VRFs, route distinguishers, BGP-VPN sessions. The orchestrator above consumes L3SM or its own CFS equivalent; the controller consumes L3NM or a vendor-native model. Conflating the two collapses the abstraction, and the catalog ends up holding both kinds of detail. OpenConfig adds operator-driven device-level models, used widely for telemetry and configuration and less for services.
Live topology
- BGP-LS carries IGP topology, traffic-engineering attributes and segment routing labels into the controller. A read-only feed of network state.
- PCEP runs between the path computation element and headend routers, for central path selection that knows about existing tunnels.
- SR-PCE computes label stacks for traffic-engineered paths and pushes them to ingress PEs.
Live state is what distinguishes a domain controller from an EMS. It also sets a rule for the catalog: topology is a runtime concern. A service model that says "use SR path X" is a deployment artefact. The right shape is "deliver this service with transport class gold between these endpoints", and the controller picks the path at fulfilment time.
Who owns what
| Entity | System of Record | System of Engagement | System of Reference | Notes |
|---|---|---|---|---|
| L3VPN / L2VPN service definition (CFS) | Service Catalog | — | Domain controller | Holds transport class and endpoints, never a path. |
| Network-level service instance (VRFs, RDs, route targets) | Domain controller | — | Resource Inventory | Rendered from L3NM or a vendor-native model; inventory should reconcile from it. |
| Live topology and TE state | Network (via BGP-LS) | Path computation element | Domain controller | Runtime state. Never copied into the catalog. |
| Devices, ports, links and logical interfaces | Resource Inventory | — | Domain controller | TMF639. Reconciled from the controller and the network. |
In practice
No operator runs a single-vendor IP estate, so multi-vendor support is a hard requirement. The common controllers take different stances: Cisco NSO reaches other vendors through its NED catalog (vendor- and version-specific driver packages); Nokia NSP is Nokia-first with multi-vendor support; Juniper Paragon is strongest in path computation; Blue Planet is multi-vendor by design with deep optical heritage; ONAP SDN-C is open source, with uneven production maturity.
Driver-based reach is both strength and cost. Every vendor software upgrade may need a driver upgrade with its own test cycle. Treating drivers as zero-cost is a common transformation budgeting mistake.
The trade-off
What it solves
Standard service models and a multi-vendor controller let services be ordered by intent (sites, class, SLA) while vendors, paths and devices change underneath.
What it does not solve
- Uneven model implementation. "Supports L3NM" still needs integration testing per vendor.
- Legacy transport that has no model-driven interface (see 9.4).
- Catalogs that already hold paths or vendor detail; those have to be remodelled.
What it adds
- Driver and adapter maintenance on every vendor software upgrade.
- Two model layers (service and network) to keep separate and mapped.
- A live-state controller becomes a dependency of fulfilment and assurance.
Architecture decision
What should the controller expose to the orchestrator above?
Standard IETF network models (L3NM, L2NM)
Choose when the estate is multi-vendor, the controller implements the models well, and portability between controllers matters.
Controller-native service models
Choose when one controller is strategic, the standard models do not cover needed features, and lock-in is accepted knowingly.
Standard models with documented extensions
Choose when most services fit the standard and a few features need vendor augments, kept out of the catalog.
Breaks if vendor-flavoured service models leak into the cross-domain layer or the catalog. A vendor change then means rework all the way up to the product.
Takeaways
Key Takeaways
- IP/MPLS, transport and optical form one orchestration domain because their service models compose.
- IETF L3SM, L3NM, L2NM and L1CSM give this domain standard service and network models that other OSS domains lack.
- Keep the service model (what the customer sees) separate from the network model (what the network does).
- Topology is runtime state from BGP-LS and PCEP; the catalog names a transport class, never a path.
- Multi-vendor reach through drivers is real and has a maintenance cost on every upgrade.
Check yourself
A product manager wants the L3VPN catalog entry to specify the SR path between two data centres for a premium tier. How do you respond?Show answerHide
Model the premium tier as a transport class with SLA attributes. The controller selects the path at fulfilment from live topology. A hard-coded path breaks the first time the topology changes.
A controller vendor says it supports L3NM. What do you still need to check?Show answerHide
Which parts of the model it implements and where it deviates, tested against your own services and device vendors. Coverage is proven by integration testing.
Why does a new two-site L3VPN usually change only two routers in a large network?Show answerHide
Per-customer state lives on the PEs (VRFs, BGP-VPN, QoS). P routers and route reflectors carry no per-customer configuration; the core is built capacity.