Research On Multi-Task Offloading Optimization Strategy Of Edge Computing Based On Graph Dependency

Recently,data generated by the Internet shows an explosive growth trend,which puts forward higher requirements for the development of communication network.Therefore,moving edge computing came into being.Mobile edge computing sinks the data processing process to the edge of the network,closer to the source of data generation.So mobile edge computing can reduce latency and improve user experience while relieving the pressure on the core network.But offloading tasks to edge servers can cause additional communication latency,which can take a toll on performance.Therefore,it is crucial to improve system performance to formulate an appropriate end-edge-cloud cooperative offloading strategy based on communication and computing information to reduce the average latency and energy consumption of all users in the network.By considering the topology of applications and channel interference between users,an Actor-Critic algorithm based on two-layer embeddings is proposed to solve the offloading problem of structured applications.It is named the ACED algorithm.Instead of binary offloading and simply scaling partial offloading problems,this thesis considers the dependencies between the tasks and models these structured applications as Directed Acyclic Graphs(DAGs).In addition,local computing,edge computing and cloud computing models are designed to formalize the optimization objective based on energy time cost(ETC).This thesis models the DAG tasks offloading problem as a Markov decision process.The state of each decision time is composed of user device state,edge device state and the information of the current task to be decided.The action is task offloading decision,and the immediate reward is obtained by the negative increment of ETC.To solve this problem,this thesis proposes an Actor-Critic algorithm based on two-layer embeddings.In order to capture the topology of the application,we use a graph convolutional network(GCN)to embed the task.In addition,a multi-layer perceptron is used to embed devices’ states in the edge computing system.These two embedding vectors input into the deep reinforcement learning(DRL)model can output the offloading strategy of the current task to be decided.Above all,in this thesis,an Actor-Critic algorithm based on two-layer embeddings is proposed to solve the problem of offloading structured applications generated by user devices in heterogeneous edge computing system.Experiments show that the ACED algorithm has advantages in reducing the average ETC compared with the traditional task offloading and scheduling algorithms.