SODA: An Adaptive Bitrate Controller for Consistent High-Quality Video Streaming

Title:SODA: An Adaptive Bitrate Controller for Consistent High-Quality Video Streaming

Authors :Tianyu Chen(University of Massachusetts Amherst);Yiheng Lin, Nicolas Christianson(California Institute of Technology); Zahaib Akhtar(Amazon Prime Video/NCSU); Sharath Dharmaji(Amazon Prime Video); Mohammad Hajiesmaili(University of Massachusetts Amherst); Adam Wierman(California Institute of Technology); Ramesh K. Sitaraman(University of Massachusetts Amherst)

Speaker:Tianyu Chen(University of Massachusetts Amherst)

Scribe:Xuanhao Liu(Xiamen University)

Introduction
The primary objective of adaptive bitrate (ABR) streaming is to enhance users’ quality of experience (QoE) by dynamically adjusting the video bitrate in response to changing network conditions. However, frequent bitrate switching may be frustrating especially during live streaming when buffer lengths are short. Ideal ABR controller seeks to push the trade-off boundary and optimize all three QoE components simultaneously: video quality, rebuffering and bitrates witching. But existing ABR controller suffer from shortcomings of their own:Model predictive controllers are hard to deploy at scale, Existing ABR controllers naively reduce bitrate switching at the expense of low video quality or more rebuffering, and Variance in network conditions or throughput prediction errors are not well tolerated by existing controllers.This paper propose a practical smoothness optimized dynamic adaptive(SODA) controller to addresses these problems while remaining deployable.

Key idea and contribution :
Given the design gaps, opportunities, and requirements, it’s important to design a theoretically sound adaptive bitrate streaming (ABR) controller that minimizes bitrate switching without compromising video quality or increasing rebuffering time, thus providing a smooth viewing experience.This paper deviate from the conventional segment-based ABR formulation and derive theoretical insights from a time-based ABR formulation. Their time-based ABR formulation treats a video stream as a continuous flow rather than a discrete sequence of segments. The objective is to minimize the overall cost given as a linear combination of the three QoE components:
image
w_n denote the average throughput during the 𝑛 th time interval, r_n the selected bitrate for that time interval, and x_n the buffer level immediately after that time interval,v( r_n) is the distortion cost, 𝑏(x_n) is the buffer cost, 𝑐(r_n, r_(n-1)) is the switching cost.
Inspired by the model predictive control framework, SODA selects a bitrate for each time interval by optimizing over the next 𝐾 time intervals and then committing to the bitrate decision for the immediate next time interval, i.e., minimizing:

Evaluation
Figure shows the mean QoE scores, utilities, rebuffering ratios and switching rates of SODA and baseline ABR controllers under each network dataset. The Puffer dataset is split into four quarters according the throughput variance (Q1 being lowest while Q4 being highest). SODA has consistently higher mean QoE scores and lower switching rates than all baseline ABR controllers under all network conditions. (Error bars represent 95% confidence intervals.) More importantly, SODA is the only ABR controller that achieves low mean rebuffering ratio and switching rate simultaneously, which translates to superior smoothness of adaptive streaming.


This result is significant because It demonstrates that SODA offers higher performance, which can significantly enhance our video watching experience, making it highly valuable.

Q :why don’t you care comparing with the RL based algorithm in your simulation?

A : Yes, thanks for the question. We actually do compare with RL based algorithm. Let me go to the empirical evaluation figure. Yes, you can see that our last baseline is compared with a reinforcement learning-based algorithm on the same infrastructure.

Q: the comparisons in the next cause the QoE maybe varies of the function.

A: Yes, that’s why we not only show QoE in our paper, but we actually show the individual components. So you’re free to examine the individual QoE components and compare our algorithm SODA with each of the baseline based on the individual components at the QoE.

Personal thoughts
This paper propose a smoothness-optimized dynamic adaptive(SODA) controller that addresses this issue in a theoretically sound way. it is readily deployable in a wide range of production environments. SODA significantly reduced bitrate switching by up to 88.8% compared to a fine-tuned production baseline. SODA’s novel time-based ABR formulation and theoretical insights shed new light on how to achieve consistent high-quality video streaming. It plays an important role in both the scientific community and people’s daily lives.