In-band SDN with multiple controllers faces challenges in coordination and state growth. Periplus addresses this through border-switch forwarding graphs, keeping state local.
The first bottleneck of in-band SDN is not the traffic control itself, but its evolution as the network grows. When a single controller can no longer handle the load, the network is divided into domains. However, a new problem arises: multi-controller coordination. Controllers must discover each other and exchange state over the same channels (in-band), without bloating the flow tables on each switch. In classical approaches, this breaks down at scale: each new controller adds state to all switches. As a result, the benefits of partitioning are nearly lost—it’s not latency that increases, but state overhead.
Periplus opts for a more localized strategy. The core idea is to limit forwarding state to only the border nodes (border switches). This is achieved using the Controller Advertisement (C-Adv) mechanism and incremental route assembly between controllers. Instead of establishing a complete path through the entire network, each domain adds only its segment to the forwarding graph. This is similar to segment routing, but with an important distinction: the graph is not fully defined at the input but is “built out” at the domain boundaries. The trade-off here is clear: more logic on border switches, but a sharp reduction in pressure on interior switches.
The implementation relies on several key mechanisms:
- C-Adv messages serve a dual role: discovery and liveness. They propagate through controlled flooding without storms.
- Each controller maintains a soft-state neighbor table with timers (alive counter).
- To prevent loops, logic similar to BGP AS_PATH is employed.
- Routes are constructed via Dijkstra within the domain, but only to the nearest border switch.
The forwarding graph is encoded in packets through NSH (Network Service Header) using Open vSwitch extensions. This is significant: no firmware modification is required. However, there is a dependency on the capabilities of OVS and its timing. In the experiment, a patch was even needed to speed up reconnection—otherwise, the controller’s behavior is masked by implementation-level delays.
A separate aspect is fault tolerance. If a controller fails, the switch does not wait for an explicit signal. Its flow entries simply expire (hard timeout), and it returns to a bootstrap state. It then reconnects to any available controller via C-Adv or ARP. This provides a simple recovery model without centralized coordination. The trade-off is a temporary loss of control and reliance on timers.
Results show that the architectural solution works precisely due to the locality of state:
- In a topology with 96 switches and up to 5 controllers, bootstrap takes about 10 seconds after partitioning.
- The time to discover controllers is a few seconds, depending on connectivity rather than their number.
- Recovery after a controller failure takes about 10 seconds (in patched OVS).
- Most importantly: the number of rules on the switch is determined by its role (root, border, interior), not the size of the network.
Interior switches maintain a constant number of rules regardless of the number of controllers. Additional state appears only on border switches and linearly depends on the number of neighboring domains, not the entire network. This is the key effect: scaling through state isolation.
It is important to note the limitations:
- There is no strict consistency model between controllers. This is left to the application level.
- Load balancing between controllers is primitive (first-come-first-served).
- All results were obtained in Mininet; behavior on a real network may differ.
The industry has long discussed that in-band SDN is a trade-off between cost and manageability. Periplus makes this trade-off more manageable: it does not eliminate complexity but localizes it. Instead of fighting against state growth globally, the system confines it to the geography of domains.
This is a pragmatic step. It is not a universal solution, but a working model for networks where a dedicated control plane is not feasible and where predictability is more important than perfect consistency.
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