车载边缘计算中推理任务的实时调度策略  

Real-time scheduling strategy for reasoning tasks in vehicle edge computing

作　　者：陈乔鑫 卢宇[2] 林兵 王素云 邵浚 CHEN Qiaoxin;LU Yu;LIN Bing;WANG Suyun;SHAO Xun(College of Physics and Energy,Fujian Normal University,Fuzhou 350117,China;Concord University College,Fujian Normal University,Fuzhou 350117,China;School of Electronics Engineering and Computer Science,Peking University,Beijing 100871,China)

机构地区：[1]福建师范大学物理与能源学院,福建福州350117 [2]福建师范大学协和学院,福建福州350117 [3]北京大学信息科学技术学院,北京100871

出　　处：《计算机集成制造系统》2022年第10期3295-3303,共9页Computer Integrated Manufacturing Systems

基　　金：国家重点研发计划资助项目(2018YFB1004800);国家自然科学基金资助项目(61672159,41801324,61972165);福建省高校产学合作资助项目(2021H6026);福建省高校产学合作资助项目(2022H6024);福建省自然科学基金资助项目(2019J01286,2019J01244,2018J01619);福建省教育厅中青年教师教育科研资助项目(JT180098);福建省社会科学规划青年资助项目(FJ2020C025)。

摘　　要：为了给存在依赖关系的车载任务提供合理的实时调度方案,以使移动边缘计算通过将车载任务调度到边缘节点处理来有效减少执行时间,设计了一种车载边缘计算中推理任务实时调度策略。将自动驾驶应用的推理过程抽象为基于有向无环图的推理任务模型;利用任务优先级评价方法界定推理任务执行顺序,继而基于深度Q学习算法为推理任务选择合适的资源节点,根据资源分配结果执行相应的推理任务。实验结果表明,该调度策略能够有效减少推理任务处理时延,且收敛性与实时响应速度更佳。To provide a reasonable real-time scheduling scheme for vehicle tasks with dependencies so that Mobile Edge Computing(MEC) can schedule tasks to be processed in edge nodes for reducing execution time effectively, a real-time scheduling strategy for reasoning tasks in vehicle edge computing was designed. The reasoning process of applications was abstracted as a model based on directed acyclic graphs;the execution order of tasks was defined according to the priority evaluation method, and the appropriate nodes for the corresponding tasks were chosen by Deep Q-learning(DQN). The reasoning tasks were performed according to the results of resource allocation. Experimental results showed that the proposed strategy could effectively reduce the processing delay of reasoning tasks. Compared with the classic algorithms, it had better performance in convergence and response rate.

关 键 词：边缘计算 自动驾驶 强化学习 计算卸载 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]