Towards Offering Performance Guarantees in Software-Defined Networks

Software-Defined Networking (SDN) promises to bring the long-sought flexibility to networks by decoupling packet handling (data plane) and network control (control plane). SDN's logically central control and poor performance of early SDN controller prototypes raised concerns about its performance and scalability. Moreover, the simple hardware-based datapath that legacy SDN assumed is a major obstacle to implement complex policies without significant performance penalty. In this thesis, I show that SDN could offer performance guarantees without sacrificing flexibility. First, I present NOX-MT and HyperFlow that are controller designs that demonstrate the performance and scalability of that logically central control could offer. NOX-MT improves NOX's (the legacy SDN controller) performance by an order of magnitude and can handle $1.6M$ events per second with $8$ cores. HyperFlow transparently captures and replays state altering events to construct an eventually consistent global view across controllers. HyperFlow's prototype can synchronize 1500 events across the network in a single RTT. Second, I propose an alternative design that embraces the notion of edge-core distinction and software at the edge to free the network core from implementing complex policies and introduce the much-needed flexibility in data plane at the edge. Software datapaths are increasingly common in today's networks but they introduce a new challenge: Performance guarantees that a software datapath provides in isolation no longer holds when it is consolidated with other software datapaths. I show that performance degradation under consolidation can be directly controlled (with less than $3\%$ error) by isolating accesses to the last-level processor cache. Finally, I present a formulation of the service-level agreement enforcement problem the solution to which is a placement of software datapaths and their cache allocations that guarantees performance objectives.Ph.D