Heavy Networking 489: Is BBR Too Unfair An Algorithm For The Internet?




Heavy Networking from Packet Pushers show

Summary: BBR is a congestion control algorithm (CCA) designed for low latency. Backed by Google, BBR’s use is growing on the Internet, including as the default CCA for YouTube. However, an element in BBR’s design means that the 1.0 version of the algorithm crowds out other commonly used CCAs, such as Cubic and Reno. As today’s guests discovered in their research, when BBR and Cubic flows share a link, a single BBR flow will take up as much as 40% of the link’s bandwidth, leaving 16 Cubic flows to divide the rest among themselves. This leads to questions about how BBR will interact with the legacy algorithms that are the status-quo today. Our guests for today’s show are Ranysha Ware, a Ph.D. student in Computer Science at Carnegie Mellon University; and Justine Sherry, Assistant Professor of Computer Science at Carnegie Mellon. Ranysha recently presented her findings on BBR at the Internet Measurement Conference in October 2019 in the Netherlands. Ranysha and Justine come on the podcast to discuss: * How BBR works, including differences with legacy loss-based CCAs * Research comparing BBR, Cubic, and Reno * Why BBR behaves the way it does * Potential impacts for the Internet * Using harm, rather than fairness, as a measure for new algorithms * More Ranysha has also created a Congestion Control Evaluation Survey for network engineers and operators to get feedback on a proposed testbed for evaluating the impact of CCAs on applications and networks. She’d appreciate your input! Sponsor: Tufin Tufin has pioneered a policy-based approach to network security management using automation and analytics. As a result, you can make network changes in minutes instead of days, reliably and securely. Tufin. The Security Policy Company. Get details at www.tufin.com. Sponsor: ThousandEyes ThousandEyes gives you performance visibility from every user to every app over any network, both internal and external, so you can migrate to the cloud, troubleshoot faster and deliver exceptional user experiences. Sign up for a free account at thousandeyes.com/packetpushers and choose a free ThousandEyes t-shirt. Show Links: Modeling BBR’s Interactions withLoss-Based Congestion Control (PDF) Ranysha Ware – Carnegie Mellon University Congestion Control Evaluation Survey Ranysha Ware on Twitter Dr. Justine Sherry Justine Sherry on Twitter BBR Congestion Control Work at Google – IETF Employing QUIC Protocol to Optimize Uber’s App Performance – Uber Engineering Blog