Seminar: Building Smarter Networks Using In-Network Computing

Abstract: High-speed programmable switches allow network operators to run succinct, customized algorithms to process individual packets, providing us a prime opportunity to improve network performance and security. In my research, I adopt approximation techniques to design algorithms that fit the switches' strict memory and computational constraints, allowing more granular network measurement and real-time, closed-loop control. I then deploy these algorithms in real-world campus and carrier networks to improve user experience. In this talk, I will first introduce BeauCoup, a framework for simultaneously executing a large number of traffic measurement queries without exceeding the memory access constraints of the switch hardware. BeauCoup utilizes coupon collectors to approximately answer each count-distinct query using less than one memory access per packet on average. Second, I will present ConQuest, an algorithm that approximately analyzes the switch's queuing buffer and identifies bursty flows. ConQuest enables the switch to react to bursts in real time and protect other user traffic from suffering high delay or packet loss.