Evaluating SDN Controller Failover with Direct Replication and Shared Data Store Architectures

Abstract

Software-Defined Networking (SDN) relies on software controllers to program data-plane forwarding, so controller availability is critical: the controller is a single point of failure where High Availability (HA) state synchronization trades recovery speed against synchronization overhead. This paper empirically evaluates two HA architectures in Ryu and Mininet: RedisHA, a shared data store architecture holding controller state in an external Redis store managed via VRRP, and HTTP-HA, a lightweight direct replication architecture. Under forced failover events, HTTP-HA substantially reduces latency, achieving a 1.73 s average control-plane recovery time (TCP) and limiting packet loss to 4.16%. Conversely, Redis-HA exhibits a 5.38 s TCP and higher latency spikes (0.61 ms average RTT) due to external database round-trips and virtual-IP takeover. We conclude that HTTP-HA suits dynamic edge deployments, while Redis-HA remains useful when an external shared state store is preferred.