Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mobile Edge Computing Networks: Online Low-Latency and Fresh Service Provisioning

Yuhan Yi, Guanglin Zhang, Hai Jiang

TL;DR

The paper tackles online provisioning of low-latency and fresh edge services in MEC by jointly optimizing service caching, task offloading, and resource allocation under AoI constraints. It introduces a Lyapunov-based framework that decouples the original NP-hard, long-horizon problem into per-slot subproblems, solved via P1(a) (convex bandwidth allocation) and P1(b) (AoI-aware decisions). A novel online integrated optimization–DRL (OIODRL) approach then combines a QCQP+SDR optimization stage to generate rough caching/downloading guidance with a Clip-PPO2 learning stage to finalize offloading, caching, and CPU allocations, achieving near-optimal performance with strong stability guarantees (gap bounded by $B/V$). Extensive simulations show OIODRL matching OPTIMAL closely and outperforming several DRL and heuristic baselines, validating the method’s practical impact for online MEC provisioning with freshness constraints.

Abstract

Edge service caching can significantly mitigate latency and reduce communication and computing overhead by fetching and initializing services (applications) from clouds. The freshness of cached service data is critical when providing satisfactory services to users, but has been overlooked in existing research efforts. In this paper, we study the online low-latency and fresh service provisioning in mobile edge computing (MEC) networks. Specifically, we jointly optimize the service caching, task offloading, and resource allocation without knowledge of future system information, which is formulated as a joint online long-term optimization problem. This problem is NP-hard. To solve the problem, we design a Lyapunov-based online framework that decouples the problem at temporal level into a series of per-time-slot subproblems. For each subproblem, we propose an online integrated optimization-deep reinforcement learning (OIODRL) method, which contains an optimization stage including a quadratically constrained quadratic program (QCQP) transformation and a semidefinite relaxation (SDR) method, and a learning stage including a deep reinforcement learning (DRL) algorithm. Extensive simulations show that the proposed OIODRL method achieves a near-optimal solution and outperforms other benchmark methods.

Mobile Edge Computing Networks: Online Low-Latency and Fresh Service Provisioning

TL;DR

The paper tackles online provisioning of low-latency and fresh edge services in MEC by jointly optimizing service caching, task offloading, and resource allocation under AoI constraints. It introduces a Lyapunov-based framework that decouples the original NP-hard, long-horizon problem into per-slot subproblems, solved via P1(a) (convex bandwidth allocation) and P1(b) (AoI-aware decisions). A novel online integrated optimization–DRL (OIODRL) approach then combines a QCQP+SDR optimization stage to generate rough caching/downloading guidance with a Clip-PPO2 learning stage to finalize offloading, caching, and CPU allocations, achieving near-optimal performance with strong stability guarantees (gap bounded by ). Extensive simulations show OIODRL matching OPTIMAL closely and outperforming several DRL and heuristic baselines, validating the method’s practical impact for online MEC provisioning with freshness constraints.

Abstract

Edge service caching can significantly mitigate latency and reduce communication and computing overhead by fetching and initializing services (applications) from clouds. The freshness of cached service data is critical when providing satisfactory services to users, but has been overlooked in existing research efforts. In this paper, we study the online low-latency and fresh service provisioning in mobile edge computing (MEC) networks. Specifically, we jointly optimize the service caching, task offloading, and resource allocation without knowledge of future system information, which is formulated as a joint online long-term optimization problem. This problem is NP-hard. To solve the problem, we design a Lyapunov-based online framework that decouples the problem at temporal level into a series of per-time-slot subproblems. For each subproblem, we propose an online integrated optimization-deep reinforcement learning (OIODRL) method, which contains an optimization stage including a quadratically constrained quadratic program (QCQP) transformation and a semidefinite relaxation (SDR) method, and a learning stage including a deep reinforcement learning (DRL) algorithm. Extensive simulations show that the proposed OIODRL method achieves a near-optimal solution and outperforms other benchmark methods.
Paper Structure (25 sections, 2 theorems, 52 equations, 7 figures)

This paper contains 25 sections, 2 theorems, 52 equations, 7 figures.

Table of Contents

  1. Introduction
  2. System Model and Problem Formulation
  3. Offloading, Downloading, and Caching decisions
  4. Delay Model
  5. Processing locally at the ES
  6. Offloading to the CS
  7. Age of Information Model
  8. Problem Formulation
  9. Problem Decoupling with Lyapunov-Based Online Framework
  10. Problem Simplification
  11. Lyapunov-Based Online Framework for Problem $\mathcal{P}1(b)$
  12. Online Integrated Optimization-DRL Method for Problem $\mathcal{P}2$
  13. Optimization Stage: QCQP Transformation and SDR
  14. Problem Simplification
  15. QCQP Transformation
  16. ...and 10 more sections

Key Result

Theorem 1

Denote the utility obtained by solving problem $\mathcal{P}2$ for time slot $t$ as $U^*(t)$ and let $U^*\triangleq\frac{1}{T}\sum_{t=1}^T\mathbb{E}[U^*(t)]$. The upper bound of the gap between the optimal long-term average utility of problem $\mathcal{P}$1(b) (denoted $U^\text{opt}$) and $U^*$ is which shows that the gap can be arbitrarily small by choosing the $V$ value. With the solution of pro

Figures (7)

  • Figure 1: System model illustration.
  • Figure 2: Learning curves with different DRL algorithms.
  • Figure 3: The performance of utility with different methods.
  • Figure 4: The impact of $V$ on the reward and utility.
  • Figure 5: The impact of $F$ on the utility.
  • ...and 2 more figures

Theorems & Definitions (4)

  • Theorem 1
  • proof
  • Lemma 2
  • proof