Comparing Link Sharing and Flow Completion Time in Traditional and Learning-based TCP

Congestion control design in TCP has primarily focused on maximizing throughput, reducing delay, or minimizing packet loss. Such has been the case in the surge of TCP approaches using machine and deep learning. However, flow completion time, average throughput, and fairness index are the key performance indicators more noticeable to users and used for applications, and thus must be evaluated. We theorize, that an ideal congestion control scheme would have a small average flow completion time, and high average throughput and high fairness index. We aim to analyze the performance of a wide-variety of congestion control schemes to determine the importance of these metrics in designing a congestion control scheme. With this objective, we propose a modified reinforcement learning version of TCP; RL-TCP+, to demonstrate how flow completion time can be minimized and to evaluate it’s impact on bandwidth sharing. Through extensive experimentation, we show that greater link-sharing and fairness do not always result in lower flow completion time, and that flow-prioritization could prove beneficial in certain scenarios.