Research On Function Module Deployment Method Based On Graph Neural Network

With the rapid development of the Internet and the updating of network infrastructure,the function module that provides services for network requests in communication networks become increasingly complex,both in their structure and in the organizational structure between them and other modules.In the face of diverse function module and the inherent graph structure in communication networks,determining how to deploy function modules to suitable hardware devices to improve network system performance is difficult.Optimization of function module deployment usually requires a combination of their inherent topology information and network topology information.When dealing with complex topologies composed of several functional modules,most of the existing research is difficult to popularize in engineering because of the absence of full use of these graph structure information or the need to use a lot of domain knowledge in applications.This thesis investigates two specific sub-problems of function module deployment:virtual network function forwarding graph deployment problem and deep learning model computation graph deployment problem.Starting from these two problems,a more universal and general solution is explored in combination with graph neural networks and deep reinforcement learning.The contributions of this thesis is as follows:Aiming at the deployment of virtual network function forwarding graph,this thesis abstracts the network service request into a virtual network function forwarding graph represented by a directed acyclic graph,and proposes a deployment strategy in dynamic scenarios.In order to effectively deal with the special graph topology and related information of network services,this thesis uses edge enhanced graph neural network and graph convolutional network to extract topology information,and combines deep reinforcement learning to propose an efficient virtual network function forwarding graph deployment algorithm.For the problem that the action space is too large when deep reinforcement learning is applied to this scene,a two-stage method is proposed in this thesis.Experimental results show that compared with the relevant research,the proposed method has a significant improvement in terms of reception rate,total cost and algorithm execution time.Aiming at the deployment problem of deep neural network computation graph in edge network,this thesis further abstracts the deep neural network computation graph from a directed acyclic graph to a activity on vertex network,and designs a deployment scheme.In order to make full use of the special structure of activity on vertex network and its critical path characteristics,this thesis uses a graph neural network based on massage-passing to fully extract the relevant features,and combines deep reinforcement learning to propose an efficient computation graph deployment algorithm.In addition,aiming at the problem that the randomness of task arrival in this scenario has a negative impact on reinforcement learning training,this thesis uses an input-driven variance reduction technique to solve this problem.Experimental results show that compared with the relevant research,the proposed method in this thesis has significantly improved the average task completion time and load.