Why does ABR fail to work in HSR
Published:
Recent study 1 observes that the unpredictability of bandwidth drop in HSR has been beyond the capacity of existing ABR algorithms and thus the major threat to the control of bitrate selection. Yet, current research on ABR design have not yet treated throughput variation in much detail.
Aside from the buffer-based ABRs 2 3, there are two types of throughput prediction method deployed in currently popular implementations:
- 1) Harmonic mean 4. It is the primitive form of prediction by assuming throughput flow is steadily smooth.
- 2) Machine learning technique. Pensieve 5 sets up offline reinforcement learning trained with network simulator with a premise that training environment can generalize to the wild. Another style of RL-based ABR is FUGU with online learning 6, which attempts to fit the network variation pattern without priori knowledge.
- 3) Exploiting side-channel information. piStream 7 employs an energy-based monitor to gather the cell-wide usage data of a given LTE base station in order to sniff the available bandwidth.
Despite their success in static or simulated dynamic situation, they might lose the edge in high-mobility settings. The first two types of ABR take throughput, download time, bitrate of a few past chunks as inputs to model the overall network condition whereas in high-mobility scene, good downlink condition of some past chunks do not necessarily indicate good or bad condition of the following chunks.
For the last one, side-channel information is helpful but non-deterministic in our case. piStream gauges the potential bandwidth in theory yet a client in HSR might not be up to achieve it owing to the poor link quality. Plus, an energy-based method is prone to error, again, for the impaired channel condition.
To summarize, it is the near-term prediction vs. fluctuating throughput contradiction that prevents current ABRs to work in high-mobility scenarios.
Jiang, Zhongbai, Changqiao Xu, Jianfeng Guan, Yang Liu, and Gabriel-Miro Muntean. 2019. “Stochastic Analysis of DASH-Based Video Service in High-Speed Railway Networks.” IEEE Transactions on Multimedia 21 (6): 1577–92. https://doi.org/10.1109/TMM.2018.2881095. ↩
Huang, Te-Yuan, Ramesh Johari, Nick McKeown, Matthew Trunnell, and Mark Watson. 2014. “A Buffer-Based Approach to Rate Adaptation: Evidence from a Large Video Streaming Service.” In Proceedings of the 2014 ACM Conference on SIGCOMM, 187–98. Chicago Illinois USA: ACM. https://doi.org/10.1145/2619239.2626296. ↩
Spiteri, Kevin, Rahul Urgaonkar, and Ramesh K. Sitaraman. 2020. “BOLA: Near-Optimal Bitrate Adaptation for Online Videos.” IEEE/ACM Transactions on Networking 28 (4): 1698–1711. https://doi.org/10.1109/TNET.2020.2996964. ↩
Yin, Xiaoqi, Abhishek Jindal, Vyas Sekar, and Bruno Sinopoli. 2015. “A Control-Theoretic Approach for Dynamic Adaptive Video Streaming over HTTP.” In Proceedings of the 2015 ACM Conference on Special Interest Group on Data Communication, 325–38. London United Kingdom: ACM. https://doi.org/10.1145/2785956.2787486. ↩
Mao, Hongzi, Ravi Netravali, and Mohammad Alizadeh. 2017. “Neural Adaptive Video Streaming with Pensieve.” In Proceedings of the Conference of the ACM Special Interest Group on Data Communication, 197–210. Los Angeles CA USA: ACM. https://doi.org/10.1145/3098822.3098843. ↩
Yan, Francis Y, James Hong, Hudson Ayers, Keyi Zhang, Chenzhi Zhu, Philip Levis, Sadjad Fouladi, and Keith Winstein. n.d. “Learning in Situ: A Randomized Experiment in Video Streaming,” 17. ↩
Xie, Xiufeng, Xinyu Zhang, Swarun Kumar, and Li Erran Li. 2015. “PiStream: Physical Layer Informed Adaptive Video Streaming over LTE.” In Proceedings of the 21st Annual International Conference on Mobile Computing and Networking, 413–25. Paris France: ACM. https://doi.org/10.1145/2789168.2790118. ↩