Research On Training Data Collection And Task Scheduling In Intelligent Edge System

The rapid development of IoT technology and the popularity of AI technology re-search have brought new opportunities for the development of edge computing.As an emerging computing paradigm,edge computing makes full use of the resources of edge devices,creating a network computing system at the edge of the network.At the same time,the emergence of edge computing provides the solution for the problem of high latency of wide area network which can't be solved in the cloud computing field for a long time,and brings low latency,fast response and better service experience for users.With the rapid development of AI technology,the intelligent edge system supported by AI algorithm is also being favored by researchers in many fields.This thesis build an intelligent edge system using the edge computing architecture and try to solve some problems that may arise in the system.The intelligent edge system is such a system based on edge computing architecture that uses the techniques and applications of artificial intelligence to provide services for users.The intelligent edge system can not only provide intelligent inference for users by using intelligent services,but also independently train the AI service model.This brings convenience to the service deployment on the edge node,and also greatly reduces the data traffic and the load pressure on the core network.This thesis consider to use multiple edge nodes for the distributed neural network training task.In the existing works,researchers often pay attention to the training itself.However,there is no concern about the training tasks and data collection that must be done before the training begins.Unlike the previous research on data collection problem,the training data is usually associated with the training task.Considering the diversification of the training tasks in the intelligent edge system and the limited resources of the edge nodes,this thesis present the problem of collecting as many training tasks and data as possible with limited resources in this thesis.This thesis propose an approximate algorithm which is called RD-TDC algorithm based on random rounding technique for this problem,and analyze the approximation ratio of RD-TDC algorithm.The RD-TDC algorithm can guarantee(1—λ)μ training tasks to be trained at least in the intelligent edge,under the condition of(?)≤λ≤1.After the training is completed,the model will be cached on the edge node for the response of the intelligent service request.Within current knowledge,the real-time requirements of most intelligent service requests are very important.Therefore,this thesis consider the impact of real-time and timeliness of task request on the system,and propose that the intelligent edge system can be viewed as a soft real-time system.An online task scheduling problem to minimize the penalty of tardiness of tasks has been proposed.Compared with the scheduling problems in other edge computing systems,the scheduling problem presented in this rhesis considers the importance of deadline of task requests and considers the limitations that the machine is usable or not.Based on the urgency of the deadline,two heuristic algorithms are proposed to solve the par-allel schduling problem which can be divided into two different phases.Simulation experiments have shown that heuristic algorithms improve system performance by 30%compared to algorithms commonly used in general systems.