Congestion Control in Highly Variable Networks

Modern applications place an enormous demand on networks to deliver high throughput and low delay. To support applications, computer networks are evolving rapidly. Several new network environments such as datacenter and wireless networks have emerged recently and become prominent. While bandwidth has been increasing steadily in these network environments, they also exhibit significant variability in network conditions. For example, the capacity of a cellular link varies with time. Deployed congestion control solutions struggle to adapt to these variations, and their performance is far from optimal in many environments: the feedback used by these schemes is often imprecise or fails to capture variations in the network conditions fast enough. To improve performance, we need accurate and timely feedback. To this end, we advocate designing separate feedback mechanisms tailored specifically to the nuances of each network environment. Understanding how conditions are varying in each environment can help us unravel what kind of information about the network conditions can improve adaption to such variations. Additionally, the feedback mechanism should be practical and only involve changes that are within the administrative and hardware constraints of the given network environment. Following this philosophy, this dissertation contributes separate high performance congestion control solutions for three prominent network environments: (1) Wireless Networks; (2) Datacenter Networks; (3) Wide-area Internet. ABC is a simple explicit congestion control protocol for network paths with wireless links. ABC adapts to variations in the link capacity quickly and accurately. Compared to deployed schemes, ABC either achieves 50% higher throughput for similar delays or 3× lower delays for similar throughput. BFC is a practical per-hop per-flow flow control architecture for datacenter networks with bursty traffic. Compared to deployed schemes, BFC responds to congestion faster, and achieves 2.3 – 60× lower tail latency for short flows and 1.6 – 5× better average completion time for long flows. Nimbus proposes a new feedback mechanism, elasticity detection, to robustly characterize the nature of cross-traffic competing a flow. Nimbus enables low delay congestion control in the Internet without any router modifications. Compared to deployed schemes, Nimbus achieves 40-50 ms lower delays in the Internet for similar throughput.