S-fabric: towards scalable and incremental SDN deployment in data centers

Abstract

Scalable and robust SDN requires the controller to be distributed. In many SDN designs, the distributed controllers are acting as replicas by forming clusters. For large-scale data centers across multiple geographically distributed locations, the controllers have to maintain a synchronized global view. These restrict themselves on single point of failure, low scalability, more communication effort, bad isolation, and rigid deployment. In this paper, we propose S-Fabric, a novel data center network design, which provides a sliced control plane and a policy-based user-defined data plane. By slicing the network through flows, and assigning a non-replica controller to each slice, S-Fabric achieves flexibility and elasticity, while ensuring isolation and separation. We leverage (but not limited to) a two-tiered spine and leaf architecture, and define forwarding rules for spine, leaf and edge switch respectively. By simplifying the flow table, S-Fabric keeps the number of forwarding rules on spine switches equal to the number of used leaf/edge ports inside a data center. By matching subnets in slices to VLANs on the edge switches, S-Fabric brings backwards compatibility to traditional data centers. S-Fabric enables an incremental deployment of SDN in traditional data centers, since it requires no SDN capability on the spine/core switches.