【作者】 李婕；

【导师】 曹洋；

【作者基本信息】 华中科技大学 ， 电子与通信工程， 2020， 硕士

【摘要】 结合了图像/视频处理、计算机视觉和通信网络的多媒体物联网有望应用于智能驾驶和视频监控等领域。针对图像/视频处理和计算机视觉等相关领域问题,深度学习成为一种主流解决方案。为满足时延敏感任务的需要,深度学习可与在网络边缘处理和分析数据的边缘计算进行结合,形成边缘智能,即在更靠近数据源的边缘节点进行深度学习模型推理。然而,边缘智能尚处发展初期,目前存在诸多问题,特别是边缘节点有限的存储资源使得其无法放置大量深度学习模型,单个边缘节点有限的计算资源使得其难以在短时间内完成计算任务等。针对这些问题,本文提出了边缘计算资源放置算法和任务分配算法,并搭建实验平台及开展验证。本文的主要研究工作如下:1、针对边缘节点存储资源不足的问题,本文特别考虑各节点被请求概率不等的情况,基于深度学习模型的请求热度和节点间可通信的网络架构提出了面向边缘智能的多轮边缘计算资源放置算法,该算法在一定程度上预测了用户请求,且考虑了边缘节点间的合作,可充分利用被请求概率较低节点的闲置存储资源。针对边缘节点计算资源不足的问题,本文基于协同处理任务思想,将多个节点组成集群,以集群为单位处理计算任务,目标为减小任务处理时间和均衡节点间负载。基于实测数据和NSGA-II(Non-dominated Sorting Genetic Algorithm-II)算法设计策略,研究子任务的分配和各个应用在多个协同的边缘节点的放置问题。基于Matlab平台的参数仿真实验表明,所提出的资源放置算法在各节点被请求概率不等,且节点容量为10GB,用户请求种类为100种时,存储资源利用率相较对比算法高70%～77%,平均时延可减少5%～21%;所提出的任务分配算法在视频总任务大小为8MB时,相较对比算法可减少45%的任务处理时间。2、为弥补现有工作仅进行参数仿真的不足,本文基于Docker和Kubernetes设计并搭建半实物实验平台,以视频处理任务作为用户请求,以基于深度学习模型创建的应用作为最小任务处理单位,在实验平台验证任务分配算法。实验结果表明,相较对比算法,本文所提出的任务分配算法可减少38%的任务处理时间和50%的节点间CPU负载差值。更多还原

【Abstract】 The multimedia Internet of Things,which combines image/video processing,computer vision,and communication networks,is expected to be applied to fields such as intelligent driving and video surveillance.For image/video processing and computer vision and other related fields,deep learning has become a mainstream solution.To meet the needs of time-sensitive tasks,deep learning can be combined with edge computing that processes and analyzes data at the edge of the network to form edge intelligence,that is,deep learning model inference is performed at edge nodes closer to the data source.However,edge intelligence is still in the early stages of development,and there are currently many problems,especially the limited storage resources of edge nodes make it impossible to place a large number of deep learning models,and the limited computing resources of a single edge node make it difficult to complete computing tasks in a short time.In response to these problems,this paper proposes edge computing resource placement algorithms and task allocation algorithms,and builds an experimental platform and conducts verification.The main research work of this paper is as follows:1.Aiming at the problem of insufficient storage resources of edge nodes,this article specifically considers the unequal probability of each node being requested.Based on the request popularity of the deep learning model and the communicable network architecture between nodes,a multi-round edge computing resource oriented to edge intelligence is proposed.Placement algorithm,this algorithm predicts user requests to a certain extent,and takes into account the cooperation between edge nodes,and can make full use of idle storage resources of nodes with a lower frequency of requests.Aiming at the problem of insufficient computing resources at edge nodes,this paper,based on the idea of collaborative processing tasks,organizes multiple nodes into clusters and processes computing tasks in clusters.The goal is to reduce task processing time and balance the load between nodes.Based on the measured data and the NSGA-II(Non-dominated Sorting Genetic Algorithm-II)algorithm,the collaborative task processing strategy is designed,and the allocation of subtasks and the placement of each application in multiple collaborative edge nodes are studied.The numerical simulation experiments based on the Matlab platform shows that the proposed resource placement algorithm has differentprobability of being requested at each node,and the node capacity is 10 GB,and when there are 100 types of user requests,the storage resource utilization rate of the proposed algorithm is 70%～77% higher than the comparison algorithm,the average delay can be reduced by 5% to 21%;when the total video task size is 8MB,the proposed task allocation algorithm can reduce the task processing time by 45% compared with the comparison algorithm.2.In order to make up for the shortcomings of only parameter simulation in existing work,this paper designs and builds a semi-physical experimental platform based on Docker and Kubernetes,takes video processing tasks as user requests,and uses applications created based on deep learning models as the smallest task processing unit.The experimental platform verifies the task allocation algorithm.The experimental results show that compared with the comparison algorithm,the task allocation algorithm proposed in this paper can reduce the task processing time by 38% and the CPU load difference between nodes by 50%.更多还原