Co-Authors: Xiaoqi Yin (CMU), Junchen Jiang (CMU), Vyas Sekar (CMU), Fuyuan Lin (ICT.CAS), Nanshu Wang (ICT/CAS), Tao Liu (iQIYI), and Bruno Sinopoli (CMU)
The volume of TCP based adaptive video streaming is growing as the queuing is becoming more critical. To achieve a good QoE, the video adaptation algorithm needs to be smart. Most algorithms start with the basic bit rate and take a long time to achieve the ideal bit rate, because they cannot predict throughput, problem that this paper addresses.
Accurate bit rate prediction is tricky, as our abilities to do so are quite limited. The authors tried some approaches, such as harmonic mean, and support vector regression, but they do not work well. This paper's main contributions are the following:
- a large scale analysis of the throughput scalability and predictability, and provides insights on how to predict throughput accurately;
- it proposes CS2P, a cross-section stateful predictor for improving bit rate selection and adaptation via throughput modeling;
- a practical implementation of CS2P video player and demonstration of QoE improvements.
To perform the throughput variability across sessions and within sessions, the authors used a proprietary dataset with more than 20 million entries collected from the iQIYI, a Chinese video provider. To capture the stateful characteristics in the evolution of the session throughput, CS2P uses hidden Markov models. Also, CS2P groups similar sessions based on the observation that many sessions share similar throughput structure and uses data from these sessions for cross-session prediction.
CS2P improves most of the QoE metrics compared to other algorithms. It performs 3.2% better than harmonic mean and 14% to buffer based solution, for instance.
Q: Link type is not used as one of the metrics. How does the algorithm work considering different link types? (4G, Wi-Fi, ...)
A: Including link type into the candidate feature set will further enhance the accuracy of CS2P, because the link type is critical for the throughput; however, we do not have this information from the content provider.
Q: When doing throughput prediction, how far ahead does CS2P can go and how does it affect the QoE?
A: CS2P looks into different horizons in the future and the improvement over other algorithms is consistent. Before the start of a session, CS2P predicts the overall downloading time for the entire video.
Q: How does the prediction vary based on where the nodes are?
A: Because CS2P relies on cross session prediction, the more sessions with similar features the more accurate it will be. Unfortunately, in our pilot experiment we do not have many volunteers.
Q: Regarding locality issue, how does it play out in a CDN with different node conditions?
A: The CDN server is in the candidate feature set. We have considered that in our experiments.