[Paper Review] Max-Weight Scheduling and Quality-Aware Streaming for Device-to-Device Video Delivery

Joongheon Kim, Andreas F. Molisch+1 more|arXiv (Cornell University)|Jun 18, 2014

Caching and Content Delivery3 references3 citations

TL;DR

This paper proposes a joint max-weight scheduling and quality-aware streaming framework for device-to-device (D2D) video delivery, integrating dynamic video quality adaptation with interference-aware link scheduling. By combining a message-passing algorithm for max-weight independent set (MWIS) scheduling with a stochastic quality control policy, the proposed mpMWIS-QP and FlashLinQ-QP algorithms achieve significantly lower stall probability and higher average video quality than conventional separated approaches, with mpMWIS-QP eliminating stalls at 8-second pre-buffering time.

ABSTRACT

We propose and analyze centralized and distributed algorithms for device-to-device video scheduling and streaming. The proposed algorithms address jointly the problems of device-to-device link scheduling and video quality adaptation in streaming. Our simulations show that the proposed algorithms significantly outperform conventional separated approaches that treat these two problems independently.

Motivation & Objective

• To address the limitations of conventional D2D video delivery systems that treat scheduling and video quality adaptation as separate problems.
• To design a centralized and distributed algorithm that jointly optimizes device-to-device link scheduling and video quality adaptation for improved quality-of-service.
• To achieve a better trade-off between video quality (PSNR) and playback stall probability in dynamic wireless environments.
• To enhance system stability and throughput by integrating queue backlog and channel state information into scheduling decisions.

Proposed method

• Formulates the D2D link scheduling problem as a max-weight independent set (MWIS) problem, where link weights are based on queue backlog and achievable rate.
• Uses a message-passing algorithm to approximate the solution to the NP-hard MWIS problem with low complexity and high performance.
• Introduces a stochastic quality adaptation policy that selects the optimal video quality mode per chunk to maximize expected quality while ensuring queue stability.
• Defines the quality control decision as a trade-off between video quality gain and the cost of increased bit rate, using a Lagrangian multiplier α to balance stall probability and quality.
• Applies the max-weight principle to distributed scheduling by setting link priorities inversely proportional to the product of rate and queue backlog.
• Models the departure rate using Shannon's capacity formula, accounting for multi-user interference in scheduled link sets.

Experimental results

Research questions

• RQ1 Can joint optimization of scheduling and video quality adaptation in D2D networks significantly improve the trade-off between video quality and playback stall probability compared to conventional separated approaches?
• RQ2 How does the performance of a centralized max-weight scheduling algorithm (mpMWIS-QP) compare to a distributed FlashLinQ-based scheduler enhanced with quality-awareness (FlashLinQ-QP)?
• RQ3 What is the impact of pre-buffering time on the number of playback stalls and average video quality in D2D video streaming?
• RQ4 To what extent does incorporating queue backlog and channel state information into scheduling improve system stability and quality under dynamic interference conditions?

Key findings

• The mpMWIS-QP algorithm achieves zero playback stalls at a pre-buffering time (PBT) of 8 seconds, outperforming FlashLinQ-QP and FlashLinQ-Q.
• FlashLinQ-QP reduces the number of stalls to zero at PBT = 9 seconds, while FlashLinQ-Q still experiences stalls even at PBT = 10 seconds.
• mpMWIS-QP delivers an average PSNR that is 0.4 dB higher than FlashLinQ-QP and 1.6 dB higher than FlashLinQ-Q under equivalent conditions.
• The proposed quality-aware streaming policy effectively balances video quality and stall probability, with performance scalable via the Lagrangian multiplier α.
• The joint optimization framework significantly improves the quality-stall trade-off, demonstrating that decoupling scheduling and quality adaptation leads to suboptimal performance.