Event Sequence-Driven Generalized and Accurate End-to-End Streaming Latency Measurement

Video streaming latency is critical for service providers for network problem diagnosing and further data analytics to provide better user experience to end users. However, there lacks a generalized approach for measuring end-to-end latency and corresponding quality of service metrics in cloud-based video streaming systems. Therefore, we proposed a new approach to measure the end-to-end latency, along with a list of valuable metrics for streaming service providers, such as video encoding/decoding latency and jitter buffer latency. Our system implements such an approach by instrumenting the video streaming infrastructure at both the server side and client side by labeling event identifiers and corresponding timestamps. Compared with traditional approaches that require manual efforts to train a deep learning model or demands ad-hoc hardware, our system significantly reduces the overhead for involving humans in the system pipeline, and also provides comprehensive streaming-related metrics. The evaluation results show the different kinds of metrics contributed to the end-to-end latency, which we think helps the community further improve their streaming systems.