Consistent high performance and flexible congestion control architecture

Congestion Control;Internet protocols;Data Transfer

Dong, Mo

关键词: Congestion Control;    Internet protocols;    Data Transfer;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/100418/DONG-DISSERTATION-2017.pdf?sequence=1&isAllowed=y

美国|英语

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

【 摘 要 】

The part of TCP software stack that controls how fast a data sender transfers packets is usually referred as congestion control, because it was originally introduced to avoid network congestion of multiple competing flows. During the recent 30 years of Internet evolution, traditional TCP congestion control architecture, though having a army of specially-engineered implementations and improvements over the original software, suffers increasingly more from surprisingly poor performance in today's complicated network conditions. We argue the traditional TCP congestion control family has little hope of achieving consistent high performance due to a fundamental architectural deficiency: hardwiring packet-level events to control responses. In this thesis, we propose Performance-oriented Congestion Control (PCC), a new congestion control architecture in which each sender continuously observes the connection between its rate control actions and empirically experienced performance, enabling it to use intelligent control algorithms to consistently adopt actions that result in high performance. We first build the above foundation of PCC architecture analytically prove the viability of this new congestion control architecture. Specifically, we show that, controversial to intuition, with certain form of utility function and a theoretically simplified rate control algorithm, selfishly competing senders converge to a fair and stable Nash Equilibrium. With this architectural and theoretical guideline, we then design and implement the first congestion control protocol in PCC family: PCC Allegro. PCC Allegro immediate demonstrates its architectural benefits with significant, often more than 10X, performance gain on a wide spectrum of challenging network conditions. With these very encouraging performance validation, we further advance PCC's architecture on both utilty function framework and the learning rate control algorithm. Taking a principled approach using online learning theory, we designed PCC Vivace with a new strictly socially concave utility function framework and a gradient-ascend based learning rate control algorithm. PCC Vivace significantly improves performance on fast-changing networks, yields better tradeoff in convergence speed and stability and better TCP friendliness comparing to PCC Allegro and other state-of-art new congestion control protocols. Moreover, PCC Vivace's expressive utility function framework can be tuned differently at different competing flows to produce predictable converged throughput ratios for each flow. This opens significant future potential for PCC Vivace in centrally control networking paradigm like Software Defined Networks (SDN). Finally, with all these research advances, we aim to push PCC architecture to production use with a a user-space tunneling proxy and successfully integration with Google's QUIC transport framework.