Deep Learning-Assisted Online Task Offloading for Latency Minimization in Heterogeneous Mobile Edge
Deep Learning-Assisted Online Task Offloading for Latency Minimization in Heterogeneous Mobile Edge
复制标题
DOI：
10.1109/tmc.2023.3285882
复制
发表时间：
2024-05
影响因子：
7.9
通讯作者：
Yu Liu;Yingling Mao;Z. Liu;Yuanyuan Yang
中科院分区：
文献类型：
--
作者：
Yu Liu;Yingling Mao;Z. Liu;Yuanyuan Yang
近年来，随着智能设备的普及，出现了许多需要高计算能力和低延迟的应用。边缘计算是支持这类应用的一个很有前途的范例。由于边缘环境的高波动性，例如，移动设备的频繁移动，不同的任务大小和时变信道条件，我们必须在飞行中做出卸载和资源管理决策。本文提出并研究了异构移动边缘环境下的在线任务卸载和资源管理问题。这个问题的目标是最小化整个系统延迟。我们证明了这个问题是np困难的。此外，传统算法需要较长的决策时间，不足以支持高波动性的应用。本文提出了一种基于深度学习的在线快速决策算法。特别地，我们为所提出的问题设计了一个离线求解器，并使用深度神经网络来模拟求解器。我们进行了大量的模拟来评估所提出的方法。结果表明，该方法的最优解和离线近似解分别比商业Gurobi求解器快约$50,000\times$50,000 x和$500\times$ 500x。此外，该方法的总体延迟接近最佳。
With the proliferation of smart devices in recent years, many applications requiring high computing capability and low latency have emerged. Edge computing is one of the promising paradigms to support such applications. Due to the high volatility of edge environments, e.g., frequent movements of mobile devices, varying task sizes, and time-variant channel conditions, we have to make the offloading and resource management decisions on the fly. This paper formulates and studies the problem of online task offloading and resource management in heterogeneous mobile edge environments. The goal of the problem is to minimize the overall system latency. We prove that the problem is NP-hard. Moreover, traditional algorithms needing long decision-making times are insufficient to support applications with high volatility. This paper proposes a deep learning-assisted online algorithm that can make fast decisions. In particular, we design an offline solver for the proposed problem and use a deep neural network to emulate the solver. We conduct extensive simulations to evaluate the proposed approach. Results show that the proposed approach is around $50,000\times$50,000× and $500\times$500× faster than the commercial Gurobi solver for the optimal solution and the proposed offline approximation solver, respectively. Moreover, the overall latency under the proposed approach is near-optimal.