Internet congestion control architecture with application-aware utilities and adaptive timescales

While the bandwidth of network connectivity has continually increased, the demand for increasingly high quality video, the needs of latency-sensitive interactive applications, and overall increased use of the network have made congestion control a persistently challenging problem. Traditional Internet congestion control primarily divides bandwidth equally among competing flows, but ignores the heterogeneous elasticity, or time sensitivity, in application requirements. In addition, we argue that with the rigid timescale in assessing the performance of different sending rates, existing protocols miss the fundamental opportunity to achieve consistently high performance that can adapt to different network conditions. In this thesis, we first propose PCC Proteus, which accommodates the elasticity in the timing requirements for some applications by developing a scavenger mode for Internet congestion control that yields to normal (primary) flows. Scavenging provides the means for applications with elastic demands to avoid utilizing excessive bandwidth at the expense of those of more immediate priority, thus benefiting users through higher Quality of Experience. Proteus is based on a utility framework, and contains novel utility functions specific to both primary and scavenger priorities. Furthermore, Proteus utilizes a modular architecture, that can flexibly combine the primary and scavenger priorities in a hybrid mode, in which the application can adjust the priority in real-time for enhanced performance. As the first work to comprehensively explore scavenger congestion control, Proteus can both effectively yield to many existing primary protocols and act as a high performance congestion control protocol in general, according to our emulation and real-world evaluation. Second, aimed at introducing adaptability to the timescales in congestion control, we design Rubato. It starts from a pacing scheme encapsulating packet bursts with multiple burst rates, and inherently supports adaptive timescales by leveraging statistical inference techniques. Depending on the network conditions, Rubato is capable of observing the network for multiple RTTs to avoid noisy rate control decisions, and allows sub-RTT timescales with the selection of different burst rates, as well. With a wide range of evaluation, Rubato is shown to achieve a better tradeoff between convergence speed and stability more consistently compared with existing protocols including Proteus. Then, to facilitate real-world adoption, we leverage Proteus's hybrid mode in the adaptive video streaming application. We also port our Proteus and Rubato implementation to QUIC, a widely deployed datapath for transport layer protocols. That experience inspires us with insights on necessary APIs for congestion control implementation.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Meng, Tong

Contributors dc:contributor

• Godfrey, Philip Brighten
• Schapira, Michael
• Mittal, Radhika
• Kravets, Robin Hillary

• Transport layer
• Congestion control
• Scavenger
• Adaptive timescale

Statement dc:rights

• Copyright 2021 Tong Meng

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/113160

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/113160