Multi-controller coordination in in-band SDN control plane becomes tractable when forwarding state stays bounded. Periplus shows how to scale without inflating switch tables.
In the in-band SDN control plane, management and data share the same network. This simplifies deployment in constrained environments but breaks down at scale. Once multiple controllers appear, a dual burden arises: controllers must discover each other and synchronize over the same network, while switches must handle controller failures. Classic approaches address this through separate spanning trees or dedicated paths. The cost is an increase in forwarding state on each switch proportional to the number of controllers. At some point, this becomes a bottleneck: tables grow faster than the network itself.
Periplus offers a more restrained approach. The core idea is to localize state at the domain boundaries. Instead of disseminating routes to all controllers across the network, the system uses Controller Advertisement (C-Adv) and partial forwarding graphs. Each border switch only adds its segment of the route to the next domain. Interior switches are completely unaware of foreign controllers. This is a trade-off: slightly more complex logic at the boundary, but a stable volume of state within the domain.
The implementation relies on existing mechanisms. It uses the Ryu SDN framework and Open vSwitch without modifications, with Nicira extensions for NSH (Network Service Header). Routing is constructed through embedding the forwarding graph directly into the packet. The controller computes the primary path (Dijkstra) and alternatives, while the switch locally fails over without contacting the controller. The same C-Adv is used for controller discovery, but with an extension: messages carry a list of known controllers and a path vector, similar to BGP AS_PATH, to avoid cycles.
Coordination between domains occurs incrementally. When one controller learns about another, it establishes a flow only to the nearest border switch. The next domain then adds its segment. As a result, the complete route is never stored in one location. This reduces memory requirements and makes the system resilient to network growth. The analogy with segment routing is appropriate, but here segments are added along the path rather than at the entry.
Controller failure is handled through soft-state. If flow entries stop updating, the switch reverts to its initial state and repeats the bootstrap. A new controller is selected on a first-come-first-served basis via anycast. Importantly, recovery does not require explicit coordination among controllers. Behavior is determined by local signals (C-Adv or ARP), which reduces latency and dependence on global state.
Experiments in Mininet demonstrate how this performs under load. In a topology with 96 switches and 5 controllers, bootstrap time stabilizes around 10 seconds when partitioned into 3+ domains. This is a consequence of reduced diameter for each domain. Discovery among controllers converges within a few seconds. Recovery after a controller failure takes about 10 seconds on average. Notably, these figures depend on the OVS implementation, and some delays are related to reconnect logic rather than architecture.
A key result is the distribution of state. Interior switches maintain a constant number of rules (~42), regardless of the number of controllers. Additional load appears only on border switches (~52 rules), where transit graphs are stored. This confirms the initial hypothesis: scaling is achieved not by optimizing algorithms, but by limiting the scope of state propagation.
In the industry, this approach appears as an evolutionary improvement. It does not require new hardware or protocols. It shifts the point where complexity is stored. Instead of “everyone knows everything,” the system becomes “everyone knows enough.” This makes the in-band SDN control plane more practical for distributed and resource-constrained networks, where a separate control-plane infrastructure is not feasible.
Limitations remain. Load balancing among controllers is still primitive. State consistency is left to a higher level. And all evaluations have been conducted in emulation, without field conditions. But the architectural idea itself—limiting the growth of forwarding state through a border-centric design—appears robust and portable.
Information source
arXiv is the largest open preprint repository (since 1991, under the auspices of Cornell), where researchers quickly post working versions of papers; the materials are publicly accessible but do not undergo full peer review, so results should be considered preliminary and, where possible, checked against updated versions or peer‑reviewed journals. arxiv.org