Enhanced Bandwidth Measurement and Robust Rate Adaptation for Low-Latency Live Streaming

Low latency live streaming (LLLS) like LL-DASH has significantly reduced the end-to-end latency via chunked transfer encoding (CTE). However, LLLS also comes with more challenges for adaptive bitrate (ABR) algorithms: (1) bandwidth measurement is non-trivial and inaccurate due to the possible idle time between chunks in CTE; (2) the various uncertainty in LLLS such as fluctuating segment size further lead to inaccurate buffer estimation, severely degrading ABR's performance. In this paper,$w$e propose AAR which comprises two modules: (1) accurate bandwidth measurement via server-side Flag parameter to identify the burst chunks within a segment, which allows for more consecutive valid HTTP chunks; (2) an LLLS tailored ABR with a novel robust objective that maximizes the minimum quality of experience (QoE) brought by the uncertainty. To obtain the minimum QoE, we propose a theorem based on the upper bound of download time estimation, which is backed up by theoretical guarantees. To derive the maximum QoE, we propose a new LLLS state evolution mechanism and apply Model Predictive Controller (MPC) to search for optimal bitrates. Extensive real-world experiments demonstrate that AAR outperforms existing baselines with 10%-80% measurement error reduction, and QoE improves by 39%-104% throughout all considered network conditions.