车联网中基于DDQN的边云协作任务卸载机制

Edge-Cloud Collaborative Task Offloading Mechanism Based on DDQN in Vehicular Networks

面对车载终端数据计算量的爆炸式增长,计算卸载是缓解车辆资源不足的有效手段。相比于单独研究云计算或边缘计算,让两者相互协作可以实现优势互补,提高系统的整体服务质量。在车联网中,制定适应环境动态性的卸载决策存在较大困难,其中任务的紧急程度也是一个不容忽视的因素。构建一个基于软件定义网络的边云协作任务卸载架构,并设计任务优先级的度量标准,将动态环境中的任务卸载决策问题建模为马尔可夫决策过程,从而最大化由时延和成本构成的任务平均效用。为了求解任务卸载决策,提出基于双深度Q网络的任务卸载决策算法以及基于优先级的资源分配方案,并设计一种卸载比例计算方法,以保障卸载的任务量能够在通信时间内上传完成的同时最小化任务处理时延。实验结果表明,相比于全部本地、全部卸载和平均分配资源3种固定的卸载算法,该算法时延和效用性能提高了2倍以上,在车辆数目适中的情况下,任务的完成比例可以稳定保持在100%。

Computation offloading is a promising scheme to alleviate the shortage of vehicle resources facing the explosive growth trend of data computation.Compared with studying cloud computing or edge computing separately,integrating with each other can realize the complementary advantages and improve the overall quality of service.In vehicular networks,a primary challenge is to make offloading decisions,which can adapt to the dynamic environment.During this process,the urgency of tasks cannot be ignored.This paper constructs a collaborative edge-cloud task offloading architecture based on Software Defined Network(SDN),where the metrics of task priority is given.The task offloading problem is then formulated as a Markov Decision Process(MDP),which aims to maximize the utility composed of delay and cost.To solve task offloading decisions,this paper puts forward a task offloading decision algorithm based on Double Deep Q Network(DDQN)and a priority-based resource allocation scheme successively.On this basis,this paper designs a method of computing offloading ratio,which aims to minimize the task processing delay while ensuring that the part of tasks can be uploaded completely within the communication time.Simulation results show that the performance of delay and utility of the proposed algorithm is more than doubled compared to other fixed offloading algorithms such as All Local,All Offloading and Allocating Resources Evenly.Under the condition of moderate numbers of vehicles,the success rate of tasks can be maintained at 100%.