D2D辅助的边缘计算任务迁移与缓存替换研究

D2D-assisted Task Offloading and Caching Replacement for Edge Computing

针对边缘网络存在的计算和带宽资源紧张导致的高时延问题,以及边缘缓存空间的有限性,本文结合设备到设备(Device-to-Device, D2D)通信技术、缓存替换策略,提出了一种D2D辅助的边缘计算任务迁移与缓存替换机制.具体地,规划了一个综合考量边缘服务器计算和带宽资源分配、任务迁移决策和缓存决策的最小化任务时延优化问题.针对该混合整数非线性优化问题,为了进一步加快求解算法取得最优处理决策,有效满足时延敏感型设备需求,结合深度确定性策略梯度算法思想,提出了一个基于优先级经验采样的任务迁移与缓存替换算法.在深度确定性策略梯度算法的网络训练基础之上,与原算法对于经验池样本随机均匀采样不同,本算法采用了一种新的样本优先级方法,即基于样本时分误差的绝对值赋予样本优先级,从而使模型网络训练改变较大的样本被采样概率增大,加速网络训练,可较快的达到稳定收敛,获取最优处理决策.最后,仿真结果表明,与其它几种基准算法相比较,该算法在网络收敛、任务时延和缓存命中率等方面具有较大优势.

Aiming at the problem of high latency caused by the shortage of computing and bandwidth resources in edge networks,and the limitation of edge cache space,this paper proposes a Device-to-Device(D2D)-assisted task offloading and caching replacement mechanism for edge computing that incorporates D2D communication technology and caching replacement strategies.Specifically,an optimization problem is formulated to minimize task latency while considering edge server computing and bandwidth resource allocations,offloading decision,and caching decision.In order to further accelerate the solution algorithm to obtain the optimal processing decision and effectively meet the demand of delay-sensitive devices,a priority experience sampling-based task offloading and caching replacement(PES-TOCR)algorithm is proposed to solve this mixed-integer nonlinear optimization problem by using the idea of deep deterministic policy gradient algorithm.Unlike the original algorithm which randomly uniformly samples experience pool samples,this algorithm uses a new sample prioritization method,with the absolute value of the time and space error of the samples to increase the probability of selecting samples with larger changes in model network training,can accelerate the network training process to reach stable convergence faster.Thereby it can quickly obtain the optimal processing decision.Finally,simulation results show that compared with several benchmark algorithms,the proposed algorithm has significant advantages in network convergence,task latency,and cache hit rate.