Learn-ing-Based On-AP TCP Performance Enhancement

Abstract

Data transmissions suffer from TCP's poor performance since the introduction of the first commercial wireless services in the 1990s. Recent years have witnessed a surge of academia and industry activities in the field of TCP performance optimization. For a TCP flow whose last hop is a wireless link, congestions in the last hop dominate its performance. We implement an integral data sampling, network monitoring, and rate control software-defined wireless networking (SDWN) system. By analysing our sampled data, we find that there exist strong relationships between congestion packet loss behaviors and the instant cross-layer network metric measurements (states). We utilize these qualitative relationships to predict future congestions in wireless links and enhance TCP performance by launch necessary rate control locally on the access points (AP) before the congestions. We also implement modeling and rate control modules on this platform. Our platform senses the instant wireless dynamic and takes actions promptly to avoid future congestions. We conduct real-world experiments to evaluate its performance. The experiment results show that our methods outperform the bottleneck bandwidth and RTT (BBR) protocol and a recently proposed protocol Vivace on throughput, delay, and jitter performance at least 16.5%, 25%, and 12.6%, respectively.