面向边缘计算的智能资源分配与计算迁移研究  

Research on Intelligent Resource Allocation and Computation Offloading Mechanism for Edge Computing

作　　者：王倩[1] 张雅文 陈思光[1,2] WANG Qian;ZHANG Yawen;CHEN Siguang(Jiangsu Key Lab of Broadband Wireless Communication and Internet of Things,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;School of Internet of Things,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学江苏省宽带无线通信和物联网重点实验室,江苏南京210003 [2]南京邮电大学物联网学院,江苏南京210003

出　　处：《中北大学学报（自然科学版）》2023年第1期48-57,78,共11页Journal of North University of China(Natural Science Edition)

基　　金：国家自然科学基金(61971235);中国博士后科学基金(面上一等资助)(2018M630590);江苏省“333高层次人才培养工程”;江苏省博士后科研资助计划(2021K501C);南京邮电大学‘1311’人才计划;江苏省研究生科研创新计划(KYCX22_1017)。

摘　　要：为了满足物联网场景中不同用户的差异化需求和提高资源利用率,以构建一个高效的边缘计算服务系统,本文提出了一种基于优先级的物联网边缘计算迁移机制。边缘节点通过感知任务的优先程度,合理分配计算资源,并提供相应的计算服务,避免时延敏感型任务因等待时间过长而导致执行失败,提高用户服务质量。通过综合考虑计算迁移决策、带宽资源和边缘节点计算资源分配,构建了一个基于优先级的任务完成总能耗最小化问题。同时,为解决上述优化问题,设计了一种基于优先级的智能资源分配与计算迁移算法。该算法通过融合深度确定性策略梯度算法的思想,构建了双重“行动者-评论家”网络架构,加快了训练过程的收敛速度;并且,为了使该算法适用于本文规划的混合整数优化问题,对连续动作输出做离散化处理,生成二进制迁移决策。仿真结果表明,本文所提机制能够获得近似贪婪算法的最优迁移与资源分配策略,与本地计算、完全迁移和DQN(Deep Q Networks)方案相比,本文提出的算法在系统总能耗方面分别平均减少约52%、 13%和7%。In order to satisfy the differentiated demands of different users in Internet of Things(IoT) scenarios, improve resource utilization, and build an efficient edge computing service system, this paper proposed a priority-based edge computation offloading mechanism for IoT. Through sensing the priority, edge nodes can provide corresponding computing services with reasonable computing resource allocation, thereby avoiding execution failure of delay-sensitive tasks due to long-time task waiting, and improving user service quality. Specifically, under the comprehensive consideration of computation offloading decision, bandwidth resource and edge node computing resources allocation, a priority-based task completion energy minimization problem was constructed. At the same time, a priority-based intelligent resource allocation and computation offloading algorithm was proposed to solve the above optimization problem. The algorithm combined the idea of deep deterministic policy gradient algorithm, and designed a dual “actor-critic” network architecture, which accelerated the convergence speed of the training process. In order to make the algorithm suitable for solving the above mixed integer optimization problem planned in this paper, a discretization operation was performed on the continuous actions to generate binary offloading decisions. Finally, the simulation results show the effectiveness of the proposed mechanism in this paper, and it can achieve the optimal offloading and resource allocation strategy which is approximate to the solution of greedy algorithm. Compared with the local computing, full offloading and DQN schemes, the proposed scheme reduces the total system energy consumption by about 52%, 13% and 7%, respectively.

关 键 词：计算迁移 边缘计算 任务优先级 深度强化学习 资源分配 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]