When Raft Meets SDN: How to Elect a Leader over a Network

This paper discusses the benefits in the operation of a Raft based SDN controller cluster, when the election of the cluster leader becomes more or less “fair. Raft is a leader based consensus algorithm, which is used by the most popular open-source SDN controllers for replicating the network state. It requires all state changes to be confirmed by the leader, thus the leader election is very crucial for the Raft performance. In case that the inter-controller communication delay is the same for all controller pairs, the election process is absolute fair, meaning that the leadership is shared equally among the controllers. In all other cases, some controllers become leaders more frequently in benefit or at a cost of the average time required for a network state update. In this paper, we model this time as a function of the leadership probabilities of the cluster controllers. We also model these probabilities as a function of the time that each controller is waiting after detecting the current leader failure and before starting its campaign. We configure different ranges for the controller waiting times, adjusting the leadership probabilities and decreasing the average response time. Our model is confirmed by testbed experimentation.