Research On Complex Task Scheduling Algorithms In Edge Computing

The rapid development of technologies such as 5G and the Internet of Things(Io T)has led to a sharp increase in the number of devices and amount of data at the edge of networks.The massive amount of data requires storage,analysis,and processing at the edge of the network.Meanwhile,to meet the Quality of Service(Qo S)requirements of delay-sensitive applications such as autonomous driving and virtual reality,the concept of edge computing has emerged.By moving computing resources and storage capabilities from traditional cloud computing centers to edge devices that are closer to the data source or data user,edge computing can satisfy users’ high-performance and low-latency computing service needs.As edge computing is applied to more and more fields,the computing tasks it faces are becoming increasingly complex.These complex tasks are often composed of a group of sub-tasks with strict real-time requirements and a sequential execution constraint,and they will continue to be executed for a period of time.The complexity of these tasks increases the difficulty of task scheduling in edge computing.Compared with cloud computing,the computing resources of edge servers are relatively limited,and a single edge server is usually unable to meet the computing needs of complex tasks.Therefore,task scheduling needs to be carried out in the edge computing system.To address the characteristics of complex tasks,this thesis models them as weighted directed acyclic graphs(DAGs),where nodes represent sub-tasks,and directed edges with weights describe the constraint relationships and data transfer relationships between sub-tasks.The scheduling problem of complex tasks is thus transformed into a DAG scheduling problem.There are two sub-problems that need to be solved when scheduling complex tasks: when a complex task arrives at the edge computing system,we need to assign an edge server to each sub-task,and we also need to decide how much computing resources to allocate to each sub-task.Since DAG scheduling is an NP-hard problem,this thesis proposes a deep reinforcement learningbased online scheduling algorithm for complex tasks.The goal is to minimize the overall task timeout rate,i.e.,to meet the deadline of complex tasks as much as possible.In order to use the information of complex tasks as input for the algorithm,it is necessary to extract features from the DAG.To this end,this thesis proposes DAG feature extraction methods based on graph convolutional neural networks(GCN)and graph attention networks(GAT).To address the problem of large-scale action space in reinforcement learning,this thesis employs a branching architecture for action selection called the Branching Dueling Q-Network(BDQ).This approach separates multiple action dimensions for independent optimization,thus reducing the complexity of the action space.Extensive simulation experiments were conducted on the proposed algorithm,and the results show that our BDQ-based complex task scheduling algorithm outperforms the baseline algorithm.Specifically,the model using GCN reduces the global task timeout rate by 4.89% compared to the baseline algorithm,while the model using GAT reduces it by 6.88%.Moreover,our algorithm also exhibits stable performance when critical parameters are varied.