Scalable distributed control plane for On-line social networks support cognitive neural computing in software defined networks

Though most of the current proposed distributed control planes maintain strong consistency among their controllers, this paper argues the strong consistency is not a prerequisite and proposes an Event Coordination System (ECS) that enables an efficient event replaying system and a distributed control plane (DisCon) using this event replaying system to construct eventually consistent global network topologies among its controllers without sacrificing scalability. Our ECS implements a novel request handling procedure that ensures a firstly received write request is firstly multi-casted, notified, and updated, so thus our DisCon can maximally ensure the same time sequence in which topology events get updated at different controllers and the constructed topologies can reflect the real network change in practice. We highlight the major mechanisms used, discuss the major causes of this eventual consistency, estimate the inconsistency window among controllers, and show how this eventual consistency does not make a big difference in supporting network applications. Experiments are conducted to evaluate our ECS and DisCon. The results show our DisCon has a larger event replay throughput and a lower event converging delay than HyperFlow, and larger flow setup rate and lower flow setup delay than most of the current distributed control planes.