Continued development of internet congestion control: Reinforcement learning and robustness testing approaches

Networks;Congestion Control

Jay, Nathan ; Godfrey ; Philip B.

关键词: Networks;    Congestion Control;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/106267/JAY-THESIS-2019.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

The relatively recent explosion of mobile traffic in the internet, combined with a near constant dependence of both mobile and desktop applications on connectivity have driven new demands on Internet infrastructure. These changes have inspired significant recent research on Internet congestion control, and the deployment of the first significant change to standard TCP-like congestion control in the past 30 years. This evidence supports two trends: first, new congestion control algorithms are needed to appropriately operate in new networks; and second, new congestion control algorithms are coming to the world, whether I design them or not. In my work, I address both trends. First, I investigate ways to automatically learn state-of-the-art congestion control algorithms using simulations of expected network conditions and a configurable reward function. Second, I created a publicly visible testbed for easy, parallel testing of new algorithms. Using this testing infrastructure, I have compared almost a dozen significant, or recent congestion control algorithms, including identifying significant oversights in two recent algorithms published at a recent top-tier networking conference.