面向硬件感知的边缘计算卸载和资源分配  

Hardware Model-Aware Joint Offloading and Resources Allocation Optimization

作　　者：郅佳琳 王楠 满毅 滕颖蕾[1] ZHI Jialin;WANG Nan;MAN Yi;TENG Yinglei(School of Electronic Engineering,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]北京邮电大学电子工程学院,北京100876

出　　处：《北京邮电大学学报》2022年第2期22-28,共7页Journal of Beijing University of Posts and Telecommunications

基　　金：国家重点研发计划项目(2021YFB3300100);国家自然科学基金项目(62171062)。

摘　　要：传统的边缘计算卸载研究并未涉及计算机硬件实现的细节,计算模型建模粗糙,优化方案精准度低。为此,提出了基于硬件实现的多用户多边缘服务器计算卸载和资源分配联合优化方案,充分考虑了计算过程硬件实现的细节,从计算机指令执行粒度出发,综合计算机输入/输出瓶颈和内存功能模块的能耗,重新建立联合优化模型,并在满足卸载任务时延要求的前提下最小化系统能耗。此外,为解决动作空间高维的问题,采用了基于深度确定性策略梯度的混合在线二部匹配算法。仿真结果表明,计算过程中的内存能耗不可忽略,且所提出的优化算法能够有效学习最优策略,对降低系统能耗具有显著作用。The traditional researches on edge computing offloading do not involve the details of computer hardware implementation. In addition, their computing model is rough, and the optimization scheme has low accuracy. To solve these problems, a hardware-based scheme that jointly optimizes computing offloading and resource allocation for multi-user multi-edge server is proposed. In particular, the details of hardware implementation in the calculation process are considered. From the perspective of the granularity of computer instruction execution, the input/output bottleneck and the energy consumption of memory functional modules are calculated first. Then, the joint optimization model is re-established. Finally, the system energy consumption is minimized on the premise of meeting the delay requirement of offloading tasks. To solve the high-dimensional problem of action space, a hybrid online two-part matching kuhn-munkras algorithm based on deep deterministic policy gradient is adopted. The simulation results show that the memory energy consumption cannot be ignored. Besides, the proposed optimization algorithm can effectively learn the optimal strategy and has a significant effect on reducing the system energy consumption.

关 键 词：移动边缘计算 计算机流水线 在线二部匹配 深度强化学习 

分 类 号：TN911.22[电子电信—通信与信息系统]