MEC-based Collaborative Computing And Resource Optimization Method

With the advent of the era of Internet of Everything,mobile intelligent devices and artificial intelligence applications have been popularized on a large scale.Traditional cloud computing exposes problems such as insufficient real-time performance,insufficient bandwidth,high energy consumption,and poor data security and privacy.Mobile Edge Computing(MEC)can provide computing services at the edge of the network near the data source,so it can make up for the shortcomings of cloud computing.However,artificial intelligence applications based on deep learning generally require a lot of computing and storage resources,while the resources of edge devices are relatively limited and cannot meet the resource requirements of intelligent applications.Secondly,in intelligent scenarios,artificial intelligence applications usually have diverse task requirements,while edge devices usually have single functions,so it is difficult to meet the diverse task requirements of intelligent applications.Aiming at these problems,the cloud server,MEC server and terminal equipment are coordinated to fully mobilize resources,which can greatly meet the needs of artificial intelligence tasks and provide users with fast,high-quality and diversified services.Therefore,this article conducts research from three scenarios of ’ cloud-edge ’ collaboration,’ edge-edge ’ collaboration,and ’ edge-end ’ collaboration.The main research work of this article is as follows :(1)Aiming at the high latency problem caused by traditional cloud computing,a single MEC server and cloud server collaborative computing and resource optimization method is proposed.The Deep Neural Network(DNN)model is improved by model compression technology to enable it to be deployed on MEC servers with limited resources.Through the model selection technology,the appropriate threshold is set and the optimal model is selected to ensure the fast and accurate execution of the computing task.Finally,the rich computing and storage resources of the cloud server are used to assist the MEC server to update the model and improve the accuracy of the edge model.The experimental results show that the average recognition time of the proposed method is reduced by 38.82%,and the accuracy can be improved by 16%,which reduces the delay while ensuring the accuracy.(2)Aiming at the problem of insufficient resources and high delay when a single MEC server performs artificial intelligence applications,a multi-MEC server collaborative computing and resource optimization method is proposed.Taking time as the measure index,a delay prediction method is proposed to predict the calculation delay and transmission delay of convolution layer and full connection layer of DNN model.Secondly,according to the computing power of each MEC server,a convolution computing task partitioning method is designed based on model cutting technology,which makes full use of the computing resources of the edge cluster and accelerates model reasoning.For overlapping area data,a partition data scheduling strategy is designed to reduce the waste of computing and communication resources caused by overlapping data.Experiments show that the proposed method can reduce the computational delay by27.5%—42.3%,which can effectively accelerate model inference.(3)Aiming at the problem that the computing power of terminal equipment is limited,facing a large number of data and computing tasks,the load is too large and cannot be processed quickly,a collaborative computing and resource optimization method for MEC server and terminal equipment is proposed.Firstly,a layer delay prediction model is proposed to predict the calculation delay of each layer of the DNN model.Then,combined with the data transmission amount of each layer of the DNN model,a potential partition point search algorithm is designed to search for potential partition points and reduce the search space of partition decision.Finally,an online task partitioning method is proposed to search for the optimal partitioning point,and the DNN model is divided between the terminal device and the MEC server to speed up the model inference and reduce the calculation delay.The experimental results show that the proposed method can reduce the delay by 15.9%—47.9%,which proves the effectiveness of the proposed method.