Mazu: Taming Latency in Software Defined Networks

Abstract

Timely interaction between an SDN controller and switches is crucial to many SDN management applications such as fast rerouting during link or switch failure and reactive path setup for latency-sensitive flows. However, our measurement study using two vendor platforms shows that the interaction latencies such as rule installation time are significant. This is due to both software implementation inefficiencies and fundamental traits of underlying hardware. To overcome the latencies and achieve responsive control, we develop Mazu, a systematic framework leveraging both the logically central view and global control in SDN, and the dissection of latencies from our measurement study. Mazu avoids the switch CPU processing tasks due to data plane packet arrivals by redirecting the packets to a fast proxy that is tasked with generating messages for the controller. Mazu presents novel controller algorithms to spread the rule updates across multiple switches, optimally ordering rules during insertion. With reduced number of rules to update per switch, and hardware-friendly ordering, rule update tasks finish much faster. Controlled simulations and testbed experiments show that our techniques can reduce the latency to update network state by almost 5X. Thus, Mazu makes SDN-based control suitably responsive for critical management applications.