Latency and Reliability-Aware Task Offloading and Resource Allocation for Mobile Edge Computing

Chen-Feng Liu, Mehdi Bennis, H. Vincent Poor

Introduction

Mobile Edge Computing (MEC) is a cornerstone of the 5G era, enabling resource-constrained devices like smartphones and IoT sensors to offload computationally intensive tasks to nearby servers. While traditional research focuses on optimizing average performance, this paper addresses the critical need for Ultra-Reliable Low-Latency Communication (URLLC). In applications such as autonomous drones or augmented reality, average latency is insufficient; even rare "tail" events—spikes in delay—can lead to catastrophic failures. This work proposes a novel framework to ensure high reliability by modeling these extreme latency events.

System Model and Methodology

The authors bridge the gap between power efficiency and extreme reliability using a decentralized resource allocation policy. The system model considers multiple User Equipments (UEs) and MEC servers, where each UE manages a task queue and must decide between local execution or offloading to one or more servers.

MEC system model with task offloading and queueing dynamics.

The methodology rests on four key pillars:

  1. Extreme Value Theory (EVT): Utilizing the Pickands–Balkema–de Haan theorem, the authors model the "tail" of the queue length distribution. This allows the system to specifically account for rare, extreme congestion events rather than just mean rates.
  2. Probabilistic Constraints: The framework imposes strict limits on the probability of queue lengths exceeding a defined threshold (e.g., a 1% violation probability).
  3. Lyapunov Stochastic Optimization: To minimize power consumption while meeting reliability constraints, the problem is transformed into a Lyapunov framework. Probabilistic constraints are handled via "virtual queues," enabling real-time decisions on local CPU frequency and transmit power.
  4. Decentralized Allocation: Each device independently determines its offloading strategy based on local states and estimated network interference, reducing central controller overhead.

Key Insights

  • The Necessity of EVT: URLLC cannot be achieved by optimizing for averages. By explicitly modeling the scale ($\sigma$) and shape ($\xi$) parameters of the Generalized Pareto Distribution (GPD), the system proactively adjusts resources before a tail event occurs.
  • Offloading Thresholds: MEC is not always the optimal choice. For light tasks or low arrival rates, the transmission delay inherent in offloading makes local computation more efficient for latency. Offloading only becomes superior once task computation intensity crosses a specific threshold.
  • Multi-Server Diminishing Returns: While accessing more servers can reduce computation time, there is a "sweet spot" (typically 2 servers). Beyond this, gains are offset by increased co-channel interference and queueing delays at the servers.

Performance Evaluation

Numerical simulations were conducted in a 100x100m indoor area with 4 MEC servers and 36 users, using task profiles ranging from simple page processing to complex video game frames.

Average end-to-end delay vs. power consumption tradeoff for different task densities.

Key results include:

  • Power-Delay Tradeoff: The Lyapunov parameter $V$ effectively balances performance. A small $V$ prioritizes minimal latency at the cost of battery life, while a large $V$ optimizes for power conservation.
  • Reliability Gains: The EVT-based approach successfully maintains delay violation probabilities as low as $10^{-4}$, significantly outperforming traditional methods that only stabilize mean queue lengths.
  • Rapid Convergence: Estimated EVT parameters converge quickly, demonstrating the system's ability to adapt to dynamic network conditions in real-time.

Conclusion and Future Work

This research provides a robust framework for achieving URLLC in MEC environments by prioritizing tail-event reliability. Future work will address the challenges of user mobility, including handovers between MEC servers, and explore more complex task dependencies where computations are linked in a task graph.

Summary updated

BibTeX
@inproceedings{liu2017latency,
  author = {Chen-Feng Liu and Mehdi Bennis and H. Vincent Poor},
  title = {{Latency and Reliability-Aware Task Offloading and Resource Allocation for Mobile Edge Computing}},
  booktitle = {Workshops @ GLOBECOM},
  year = {2017},
  doi = {10.1109/glocomw.2017.8269175},
}

Loading related content...

Figure