Service provisioning and profit maximization in network-assisted adaptive HTTP streaming

Zhisheng Yan, Cédric Westphal, Xin Wang, Chang Wen Chen

Abstract

Modern video streaming relies on client-side adaptation logic, which often leads to suboptimal performance in shared mobile network environments. This paper proposes a network-assisted framework for adaptive HTTP streaming that centralizes video rate selection within a mobile network controller. By modeling the problem as a Markov Decision Process (MDP), the system balances user Quality of Experience (QoE) with the network operator's economic goals. Simulation results show that this approach effectively resolves instability and unfairness issues while maximizing operator profit, performing near the theoretical ideal.

Introduction

Dynamic Adaptive Streaming over HTTP (DASH) is the standard for modern video services like Netflix and YouTube. Traditionally, the adaptation logic resides on the client device, which independently measures bandwidth and requests video segments at appropriate bit-rates. However, this decentralized approach is problematic in mobile cellular networks where multiple users share a single bottleneck link. Clients, being "blind" to each other's states, compete aggressively for resources, leading to rapid quality fluctuations (instability), unfair bandwidth distribution, and a lack of mechanisms for network operators to manage congestion costs or prioritize premium services.

System Architecture and Methodology

To address these challenges, the authors introduce a network-assisted adaptation framework. The core of this architecture is a centralized "Application Function" (AF) proxy situated within the mobile network.

The proposed network-assisted streaming architecture featuring a centralized Application Function (AF) proxy.

The AF proxy intercepts and modifies client requests based on a global view of the cell's resources. The decision-making process is modeled as a Markov Decision Process (MDP) designed to maximize the operator's long-term profit. The profit model incorporates:

  • Income: Revenue generated from providing stable, high-quality playback.
  • Costs: Penalties associated with video buffering, frequent quality switches (jitter), and bandwidth consumption during peak congestion.

The system state is defined by the current bit-rates and estimated available bandwidth for all users. Using Dynamic Programming (specifically Bellman value iteration), the controller determines the optimal bit-rate for the next video segment for every user simultaneously, effectively arbitrating competition and preventing the "tragedy of the commons."

Performance Evaluation

The framework was evaluated using the ns-2 network simulator. The test video was the H.264/AVC encoded "Big Buck Bunny" sequence, offered at five quality levels (95 Kbps to 798 Kbps). A four-state Markov model simulated realistic wireless channel fluctuations.

The proposed MDP-based algorithm was compared against a "Myopic" (standard client-centric) system and an "Ideal" (perfectly predictive) upper bound.

Comparison of video session profit against streaming service rate constraints (a) and service duration (b).

Key Results

  • Profit Maximization: As shown in the figures above, the proposed algorithm achieves profits nearly identical to the "Ideal" scenario, significantly outperforming the Myopic approach, especially as service constraints and durations increase.
  • Differentiated Service: The framework allows operators to tune "priority coefficients" to offer tiered services. In simulations with a high-priority user (UE1) and a regular user (UE2), the system maintained a higher average bit-rate for UE1 (303 Kbps vs 150 Kbps) with zero buffering, whereas the Myopic system caused both users to suffer ~18% buffering ratios.
  • Stability: By centralizing control, the system significantly reduced erratic bit-rate switching, providing a smoother visual experience compared to traditional DASH.

Conclusion

This research demonstrates that shifting video adaptation intelligence from the client to the network can significantly improve both user QoE and operator profitability. While the current MDP formulation faces potential scalability challenges due to state-space explosion, the use of proxy clouds or approximation techniques offers a path forward. Future work will focus on scaling the system to larger user bases and integrating more sophisticated perceptual QoE models.

Summary updated

BibTeX
@inproceedings{yan2015service,
  author = {Zhisheng Yan and Cédric Westphal and Xin Wang and Chang Wen Chen},
  title = {{Service provisioning and profit maximization in network-assisted adaptive HTTP streaming}},
  booktitle = {ICIP},
  year = {2015},
  doi = {10.1109/icip.2015.7351310},
}
Figure