Investigation On Model And Algorithm Optimization For Virtual Network Embedding

With the continuous development of 5G and 6G,more and more new technologies are used to support the corresponding functional requirements,and the original underlying infrastructure gradually cannot meet the requirements,so resulting network rigidity problem.As a key technology to overcome network rigidity problem,network virtualization has become a hot topic of research.The purpose of network virtualization is to enable more heterogeneous networks to coexist in a physical network,and the substrate network provides services for these heterogeneous virtual networks with different requirements.Network virtualization by implementing the rational distribution of substrate resources,greatly improves the resource utilization of the underlying infrastructure.However,how to effectively allocate substrate resources to the corresponding heterogeneous virtual networks is a key and important problem,which is called the Virtual Network Embedding(VNE)problem.The process of virtual network embedding is generally divided into node embedding process and link embedding process.Most traditional methods usually embed virtual nodes according to the CPU resources of nodes,and some of them combine with local topology to embed nodes.However,virtual network embedding methods of the whole network topology are almost not considered,resulting in the problem of high cost/benefit of physical network.However,due to the multiple iteration update strategy,the existing machine learning methods take too long to run,which leads to the problem of high cost/revenue of substrate network in VNE algorithm.According to these problems,this paper mainly by considering the performance parameters of runtime and cost/revenue to study the VNE problem.Benefiting from the structural characteristics of graph topological attributes,this paper first proposes the virtual network embedding algorithm BC-VNE based on core and coritivity theory.The core and coritivity theory can effectively find the important node sets in the graph structure,and then arrange them by the value of the eigenvector centrality to embed the important nodes first.In addition,this paper designed a node average bandwidth resource attribute,so as to increase a bandwidth constraint conditions in the nodemapping process,different from other existing algorithms which using local bandwidth resource constraints,within the scope of the proposed constraint condition is based on the global network topological attributes bandwidth resource properties,thus effectively realizing the pre-determination process of the link embedding process.The performance of the BC-VNE algorithm is verified by simulation experiments,the results show that the runtime of the BC-VNE algorithm is stable and lower,and the value of cost/revenue is reduced.These show the comprehensive performance of the BC-VNE algorithm is better than other compared algorithms.In order to solve the problem that VNE method of machine learning needs a lot of runtime,this paper proposes an embedding algorithm of Fusion Graph Convolutional Network(G-VNE)using machine learning method,but not completely using machine learning.The G-VNE algorithm uses the five node attributes which extracted from the underlying network,and combined with the topology structure of the network graph to do convolution operation to obtain the node classification result.Then in the process of virtual node embedding,the candidate node set with higher classification level is embedded first.The classification results obtained based on the training of multiple attribute characteristics of nodes can effectively integrate multiple constraints,coordinate node embedding and link embedding effectively.The performance of GVNE algorithm is verified by simulation experiments,the results show that G-VNE algorithm has more advantages in cost/revenue and network request acceptance ratio without losing too much runtime.