Arsenal: Understanding Learning-Based Wireless Video Transport via In-Depth Evaluation

Recent years have witnessed a rise of learning-based (i.e., artificial intelligence driven or AI-driven) video transport design, in order to achieve consistently high performance, even when the modern Internet is becoming increasingly heterogeneous while the applications are becoming unprecedentedly demanding (i.e., the simultaneous high-throughput and low-latency requirements of HD video telephony or intelligent remote driving). While separate evaluation using proprietary platform has shown the advantage of AI-driven algorithms over their non-AI counterparts, a systematic study is missing for directly comparing these AI-driven design under a uniform and practical platform. To bridge the gap, in this work, we first design and implement a full-fledged evaluation platform named Arsenal, which incorporates multiple state-of-the-art congestion control algorithms, most of which are AI-driven. Using Arsenal, we carry out a thorough comparative study of the algorithms, over massive traces collected from heterogeneous networks including WiFi, 4G, 3G and even the rare commercial 5G wireless networks. In particular, to enable convincing measurements for the dominated real-time video applications, we collect millions of practical video sessions in cooperation with a prevailing video service provider. The evaluation provides a handful of important observations, which are undiscovered before and have important impacts on future protocol design. Moreover, we will make the platform and algorithms open-source to enrich the research tools in the intelligent transportation community.