Abstract
This paper introduces Delay-Aware Resource Allocation (DARA), a novel non-stationary framework designed for multi-user video streaming in resource-constrained Internet-of-Things (IoT) networks. By modeling packet deadlines as time-decaying weights, DARA optimizes timeslot allocation without requiring complex, real-time packet-level metadata. Experimental evaluations using real-world surveillance data and high-fidelity simulators demonstrate that DARA significantly outperforms traditional stationary policies like Round-Robin, maintaining high video quality (PSNR) and ensuring deadline-abiding delivery even under high network load.
Introduction
Streaming high-quality video over IoT-enabled networks—such as wireless sensor networks used for surveillance—is notoriously difficult. Video data is both "heavy" (requiring high bandwidth) and "perishable" (subject to strict delay deadlines). If a packet fails to arrive before its deadline, it is rendered useless, leading to severe degradation in the reconstructed video quality.
The core challenge lies in resource allocation: how should a central router, specifically a Low Power Border Router (LPBR), decide which sensor gets to transmit in which timeslot? Traditional solutions often fall into two traps. First, they may require exact distortion impact and deadline metadata for every single packet, which creates prohibitive overhead for low-power IoT devices. Second, common solutions like Round-Robin are stationary, following fixed patterns that fail to account for the statistical reality that earlier slots are more valuable for meeting deadlines.
The DARA Framework
The authors propose the DARA approach, which shifts the focus from individual packet tracking to the statistical characteristics of the video stream. The framework is built on three primary pillars:
- Weight Distributions: Instead of tracking every packet, the system assigns a weight to each timeslot within a resource block. These weights are monotonically decreasing, representing the probability that a packet sent in that slot will meet its deadline.
- Non-Stationary Policy: Unlike stationary policies, DARA's decisions for the current slot depend on the allocations of previous slots. It utilizes a resource allocation metric that balances three factors: Distance (how far a sensor is from its target rate), Benefit (the immediate value/weight of the current slot), and Urgency (the discounted sum of remaining opportunities).
- Two-Step Optimization: The LPBR first solves a convex optimization problem to find target weighted sum rates for each user. Then, the DARA algorithm iteratively assigns slots to the user with the highest calculated metric.
The paper provides a rigorous mathematical proof that this approach is optimal for weight distributions that decrease exponentially over time, effectively treating time as a decaying asset.
Experimental Setup
The framework was evaluated using the PETS 2007 Benchmark Data, featuring real-world surveillance video from four different camera views. The videos were encoded using H.264/AVC and downsampled to fit the constraints of IoT physical layers. Simulations were conducted using the 6tisch Simulator, a high-fidelity tool for the IEEE 802.15.4e TSCH (Time Synchronized Channel Hopping) MAC layer.
As illustrated in the bitstream characteristics above, different sensors exhibit varying demands, particularly when transmitting large I-frames versus smaller P and B frames. The decaying bars represent the size of the non-expired video bitstream, highlighting the "perishable" nature of the data that DARA is designed to manage.
Results and Discussion
The authors compared DARA against three benchmarks: standard Round-Robin (RR), Rate-proportional (R-RR), and Rate/Delay-proportional (RD-RR).
- Video Quality (PSNR): In average-case scenarios with 10 sensors, DARA achieved a mean PSNR of 26.9 dB (considered good quality for surveillance), while Round-Robin plummeted to 15.2 dB (unusable quality).
- Deadline Performance: DARA was the only solution that ensured all video bitstreams were delivered in a deadline-abiding manner with acceptable quality.
- Robustness: Even in "worst-case" scenarios where sensors had synchronized high-demand I-frames, DARA maintained significantly higher utility than stationary benchmarks by utilizing its non-stationary memory to "catch up" on neglected sensors.
Conclusion
DARA demonstrates that statistical information about video sources is sufficient to achieve near-optimal resource allocation in IoT networks. By moving away from deterministic packet-level metadata and stationary policies, DARA provides a low-complexity, high-performance solution for delay-constrained video streaming. Future work may explore the impact of high-interference environments and the integration of these policies into heterogeneous networks combining different wireless standards.
Summary updated