On the Design of Practical Fault-Tolerant SDN Controllers

Abstract

The increase in the number of SDN-based deployments in production networks is triggering the need to consider fault-tolerant designs of controller architectures. Commercial SDN controller solutions incorporate fault tolerance, but there has been little discussion in the SDN community on the design of such systems and the tradeoffs involved. To fill this gap, we present a by-construction design of a fault-tolerant controller, and materialize it by proposing and formalizing a practical architecture for small to medium-sized networks. A central component of our design is a replicated shared database that stores all network state. Contrary to the more common primarybackup approaches, the proposed design guarantees a smooth transition in case of failures and avoids the need of an additional coordination service. Our preliminary results show that the performance of our solution fulfills the demands of the target networks. We hope this work to be a first step in what we consider a necessary discussion on how to build robust SDNs.

A Fault-Tolerant and Consistent SDN Controller

“…However, this approach does not consider consistency issues. On the contrary, two papers on the master-slave architecture that consider the consistency problem are [2] and [10]. Both propose an architecture that uses a shared datastore on top of the controllers, in order to keep a global common NIB.…”

“…4) The controller buffer B C provides the requested message. Previous works agree that the most effective way to detect controller failures is through a distributed datastore system ( [1], [2], and [7]). For instance, [1] mentions the use of ZooKeeper [5] to detect and react to controller failures.…”

“…Even though this is a much simpler design, the available solutions use similar tools to those used in the distributed-load case, where the performance is drastically affected. For instance, the implementation presented in [2] shows that the controller throughput may be reduced almost five times when the datastore is used regularly. This leads to the second challenge:…”

“…On the other hand, there are simpler approaches known as Master-Slave, where a single controller is in charge of all decisions. It is supported by backup controllers having a synchronized view of the network, which can take the responsibility for future decisions in the case of a failure [2], [7], [10]. Further details on the load-distributed and MasterSlave approaches will be provided in Section II-C.…”

ReCSDN: Resilient Controller for Software Defined Networks

“…Cui et al [49] performed attack detection by neural network techniques. Botelho et al [50] has replicated the sheared database of the whole network state to improve reliability.…”