Latency prediction for video frames in edge video streaming with BBR controlled UDP

With the omnipresence of high-speed internet, there is a booming market for video streaming applications. Several companies compete in this market, such as Netflix, Amazon Prime, YouTube, etc. Everyone is improving their streaming technologies to strive for a better quality of experience (QoE) for their user base. In this regard, there is a growing interest in developing user space protocols for video streaming applications requiring low latency to deliver better QoE. While conventional solutions focus on adapting streaming quality by monitoring network conditions over a long duration, I explore an alternative approach for predicting the latency of an upcoming video frame delivery by monitoring the per-packet data of the current transmission. This will provide a finer-grain control to optimize streaming quality by adjusting the video frame resolution or stream quality on a more refined per-transmitted-frame basis. This will enable a better responsive adjustment for users' QoE. For this, I designed and implemented my own user-space, acknowledgment-oriented, reliable delivery transport layer protocol based on UDP which supports flow and congestion control mechanisms. This protocol has its own network packet frame formats, protocol execution state machine, packet queuing algorithms, and customized adaptation of Google’s first version of the BBR congestion control algorithm. For evaluation purpose, I conduct data transmission experiments with controlled network conditions using Wi-Fi as an edge network environment and static camera images to emulate video streams. Using session-level and individual packet-level statistical metadata, and a quantile regression prediction model with several features, I demonstrate how much better can the per-packet timing data gained from having insight into the transport protocol predict the end-to-end latency of a message, compared to having only the outside perspective, i.e., the total transmission time from the last few messages. In conclusion, my thesis evaluation shows promising results for predicting the latency of message transmission in an edge network environment with BBR-controlled UDP.