Lynx: Queueing-theoretic Congestion Control Robust to Large Number of Flows
In this paper, we propose a novel congestion control algorithm (CCA) that can maintain a low and nearly constant buffering delay while ensuring high throughput and high throughput fairness even when the number of flows sharing the same bottleneck link increases significantly. Our proposed CCA uses methods formalized in Markovian queueing models for predicting future congestion and determining its congestion window accordingly. We name this CCA as Lynx. We evaluate Lynx through emulations and experiments over the live Internet by comparing it with other CCAs, including Copa, which is known to achieve high performance relative to other CCAs and is being increasingly used by Meta/Facebook. Results show that even when sharing the bottleneck link with many flows, compared with Copa, Lynx can improve buffering delay by up to 56% and Jain's fairness index (JFI) by up to 46% while maintaining high throughput. In addition, the live Internet experiments confirm that Lynx can lead to high performance in terms of (i) throughput, (ii) round-trip time (RTT), and (iii) throughput fairness.