Software-defined networking (SDN) offers greater flexibility than traditional distributed architectures, at the risk of the controller being a single point-of-failure. Unfortunately, existing fault-tolerance techniques, such as replicated state machine, are insufficient to ensure correct network behavior under controller failures. The challenge is that, in addition to the application state of the controllers, the switches maintain hard state that must be handled consistently. Thus, it is necessary to incorporate switch state into the system model to correctly offer a “logically centralized” controller.
We introduce Ravana, a fault-tolerant SDN controller platform that processes the control messages transactionally and exactly once (at both the controllers and the switches). Ravana maintains these guarantees in the face of both controller and switch crashes. The key insight in Ravana is that replicated state machines can be extended with lightweight switch-side mechanisms to guarantee correctness, without involving the switches in an elaborate consensus protocol. Our prototype implementation of Ravana enables unmodified controller applications to execute in a fault-tolerant fashion. Experiments show that Ravana achieves high throughput with reasonable overhead, compared to a single controller, with a failover time under 100ms.
Ravana: Controller Fault-Tolerance in Software-Defined Networking
Naga Katta, Haoyu Zhang, Michael J. Freedman, Jennifer Rexford
Proc. ACM SIGCOMM Symposium on SDN Research
(SOSR ’15) Santa Clara, CA, June 2015. [pdf]