Rate control algorithms are central to the Internet, at both the transport and application layers. At the transport layer, congestion control (CC) has a crucial impact on user experience for Internet services such as video streaming and voice-over-IP, as well as emerging Internet services such as augmented and virtual reality, Internet of Things, edge computing, and more. At the application layer, adaptive bitrate (ABR) video streaming algorithms optimize video quality in real time.  Rate control has emerged as one of the most important persistent challenges in networking. Rate control algorithms have to deal with networks that vary greatly in sizes, link capacities, latency, level of competition between connections, and more factors, all of which can change rapidly. Despite receiving only a small trickle of information about the network (one data point per packet), rate control algorithms must make millisecond-level decisions. Even recently-deployed protocols like BBR, deployed by Google and Microsoft respectively, have significant performance shortcomings.

This project seeks to develop a fundamental understanding of the use of reinforcement learning for rate control, and is applying that knowledge to improve protocol performance and functionality in ABR video and congestion control. To this end, the project produced four key technical results.  First, the project developed PCC Proteus, a "scavenger" congestion controller, which utilizes available bandwidth for non-time-sensitive workloads (such as software upgrades and backups) without disturbing time-sensitive workloads.  PCC Proteus leveraged an architecture based on online learning, and outperformed past scavenger designs (including LEDBAT).  Second, the project extended that learning-based congestion control framework to support multiple paths, significantly outperforming the widely-used Multipath TCP (MPTCP) protocol.  Third, the project developed a multipath packet steering system called DChannel that utilizes two channels -- one high bandwidth, one low latency -- simultaneously. Such channels have been standardized in 5G cellular networks, where the project's experiments show that even though the low-latency channel offers just 1% of the bandwidth of the high bandwidth channel, DChannel's use of both channels can improve web page load time and responsiveness of common mobile apps by 16-40%.  Finally, the project developed a system to use machine learning as an adversary, to perform robustness testing of ABR video and congestion control protocols.  This helps protocol designers to discover environments where their protocols perform poorly, and to fix the discovered problems.

These results can have broad impact. Congestion control protocols are used in nearly every networked application. Multiple paths or communication channels are becoming an important scenario, given the prevalence of mobile devices with multiple network interfaces and emerging 5G slicing. The project's results can broaden the applicability of these emerging technologies. Finally, the results on adversarial testing are important to improve the dependability of these core infrastructure protocols. The project trained multiple graduate and undergraduate student researchers. Results were disseminated via publications, presentations, and open-source releases of code and data.

Last Modified: 03/16/2024
Modified by: Philip B Godfrey