Research On Resource Allocation And Task Scheduling For Edge Intelligence

In recent years,the marginalization of network and computing infrastructure becomes the trend of network development and computing architecture technology.Intelligent network and system forms based on edge infrastructure are gradually formed,and the perception and computing models based on edge computing are promoted.Compared with traditional centralized networks and computing architectures,the decentralization of edge network devices will lead to localization of computing,storage,and transmission resource allocation strategies.Related services and applications are dynamically affected by user movement and interaction behaviors,making resource allocation strategies inefficient.Therefore,edge intelligence brings new challenges to network architecture,data transmission,resource allocation,model deployment,and distributed training,etc.By making full use of edge network resources and dynamically adjusting it,improving user service quality and reducing operator's operating costs are the key points to achieve high-performance large-scale edge intelligent services.This paper addresses some key issues of resource allocation and task scheduling in edge intelligence through fine-grained sensing,managing,and scheduling of edge network resources,combined with analysis of three mainstream application scenarios in edge intelligence.The main contributions of this thesis are as follows:(1)Aiming at the issue of multi-user service quality fairness in edge video transmission,a dynamic adaptive bandwidth allocation strategy based on reinforcement learning is proposed.This dissertation first analyzes the best entry point to resolve multi-user fairness in the HTTP adaptive streaming ecosystem: the network-side solution.Then we design an edge-oriented network-driven adaptive streaming media transmission framework to achieve a fair quality of experience with multiple players competing.To solve the online dynamic bandwidth allocation problem,we propose a method based on reinforcement learning,which can gradually learn and improve the fair allocation strategy through the interaction and trial and error of the video player with the environment.When the reinforcement learning algorithm training converges,the algorithm will learn a better strategy for allocating bandwidth between players to achieve a fairer multi-user quality of service experience.(2)Aiming at the cost optimization of large-scale inference services in edge inference,a multi-version model and a multi-data model adaptive inference mechanism are proposed.This work first measures the impact of model "compression" and data "compression" on inference accuracy.Different versions of the model and data can achieve similar inference accuracy.Then,we designed an inference service solution.This method satisfies the response time and accuracy of the user's requirements for the inference service.The method uses the compromise between computing resources and bandwidth resources,and adjusts the model and data version to minimize the overall service cost for large-scale inference system.We deploy different versions of the machine learning model to distributed edge servers and schedule the user's inference service requests to strategically redirect the request to the corresponding edge server and specify the user to upload data using a particular version.To solve the joint optimization problem of model deployment and task scheduling,we proposed an effective rounding-based approximation algorithm to solve the problem and provide a theoretical performance guarantee.Real experiments deployed on the Amazon Cloud show that our method can effectively reduce service costs and guarantee user response time and inference accuracy requirements.(3)Aiming at the bandwidth bottleneck of distributed training in edge intelligence,a model selection aggregation,and bandwidth-aware node selection mechanism are proposed.This work designs a decentralized federated learning framework.By using gossip protocols,we achieve global model convergence by exchanging information between the nodes.To reduce the bandwidth bottleneck during model updates aggregation,we split the model updates that need to be transmitted into multiple model segments with the same size,and each node performs local model aggregation by pulling the corresponding model segments from peers.In order to further reduce the transmission delay of the model during the synchronization phase,we designed a method of bandwidth-aware node selection,which accelerates transmission by selecting nodes with high transmission speed with probability.Experiments show that our method not only can significantly reduce the total training time,but also guarantee the convergence effect.