Research On Key Technologies Of Resource Management For Edge Intelligenc

As a new distributed computing paradigm,edge intelligence can effectively reduce the bandwidth load pressure of the core network by sinking intelligent mobile applications(e.g.,autonomous driving)from the cloud data center to the edge of the network;and resources are deployed on edge servers closer to end devices,which can further improve the execution efficiency of task processing.Resource management in an edge intelligence environment mainly includes resource allocation,resource deployment,and resource adjustment.By effectively coordinating and optimizing edge resources,end users can be provided with low-latency and high-quality intelligent applications.However,intelligent applications often rely on Deep Neural Networks(DNNs),which have the characteristics of computation-intensive and resource-intensive task processing.In addition,the resource capacity of the edge side is relatively limited,the edge server nodes are heterogeneous,and the edge network topology is dynamic and time-varying,making it difficult to guarantee the quality of service for users.Therefore,resource management in edge intelligent environments faces the following challenges: 1)For the edge training of DNN models,the heterogeneity of server nodes and the diversity of user service requirements lead to the training process with imbalance in computing nodes;2)For edge deployment of DNN models,existing methods often ignore the multi-dimensional resource requirements(processor,memory,storage,network)of DNN execution,resulting in increased deployment cost and decision complexity;3)The dynamic mobility of end devices and the randomness of edge network topology make computing tasks face scheduling out-of-order problems,resulting in degradation of task execution performance,making it difficult to meet users’ low-latency service demands.Based on the above analysis,this thesis conducts research on key technologies for edge intelligence-oriented resource management.The main research work includes:(1)In order to solve the problem of node resource imbalance in the edge training process,this thesis proposes a dynamic node resource allocation strategy for edge training,which aims to improve the training efficiency while ensuring the utilization of node resources.Specifically,for local end data upload,edge side training execution,and cloud data center aggregation model,the effective scheduling of edge server computing and communication resources is studied,and a dynamic resource allocation optimization problem is constructed.To further strengthen the load balancing of servers under edge training,a distributed collaborative training algorithm based on lightweight is designed.Furthermore,the problem of joint allocation in computing and communication resources is transformed into a Markov decision process,and a distributed dynamic resource allocation algorithm is proposed,based on the policy gradient technique,and the stability of resource allocation policy generation is ensured by pruning operations.Theoretical analysis and experimental results show that the proposed algorithm effectively reduces training delay and energy consumption while saving data transmission overhead.(2)In order to solve the problem of inaccurate decision-making in edge-side DNN model deployment,this thesis proposes a cost-aware model efficient deployment strategy under resource constraints.First,based on the processing delay and energy consumption executed by DNN,the cost and benefit optimization problem of edge server model deployment is constructed to maximize the benefits of edge side deployment.Considering the resource competition problem of multiple edge servers,the problem is modeled as an exact potential game.Furthermore,aiming at the performance degradation problem of traditional centralized decision-making control under large-scale nodes,a decentralized deployment decision-making algorithm is proposed.Realize that each edge server independently makes deployment decisions as a participant without any global information interaction.Theoretical analysis and experimental results show that the proposed method can obtain higher system benefits when deploying the model on the resource-constrained edge side.(3)In order to solve the out-of-order problem of computing task scheduling caused by the dynamic geographical location of end devices,this thesis proposes a task scheduling strategy based on popularity prediction in mobile scenarios.First,a spatiotemporal-aware popularity prediction algorithm is proposed by analyzing the context features of requests.Furthermore,based on the analysis and prediction results of future request popularity,a task scheduling algorithm that supports data caching and resource allocation is designed to improve task scheduling efficiency at the edge side and optimize processing delay.Specifically,online task scheduling,caching,and resource allocation decisions are realized by sensing the task request status of end devices,geographic location information,and computing execution status of edge servers in the time-varying environment.Theoretical analysis and experimental results show that the proposed method can achieve lower computing task processing delay in dynamic scenarios.