Research On Virtual Network Construction And Resource Management In Network Emulation Platform

With the development of network emulation platform(i.e.,Cyber Range),the test scene’s demand for high fidelity and resource utilization efficiency has increased significantly,bringing significant challenges to Cyber Range’s development.First,when constructing a multi-dimensional virtual network(VN)involving real servers,virtual machines,containers,and network simulators,it is challenging for Cyber Range to establish a new efficient network embedding model and provide a high-fidelity network emulation environment.Second,when multiple emulation nodes compete for physical resources in Cyber Range,the performance of emulation application programs will drop sharply.Host virtual resources management and cluster resource sharing become the bottleneck of physical resource utilization and emulation application performance in Cyber Range.Therefore,it is necessary to build efficient embedding and network emulation strategies for multi-dimensional VNs and propose a cluster resource collaborative optimization method based on local adaptive virtual resource management,to improve emulation fidelity and resource utilization efficiency in Cyber Range.This thesis focuses on the VN construction and resource management in Cyber Range.It mainly consists of four aspects as follows:First,during VN construction,this thesis studies a multi-dimensional VN embedding method to provide Cyber Range with deploy strategies that optimize cost and revenue.Aiming at the shortcomings of traditional methods in terms of VN scalability and diversity demands,this thesis builds a repeatable multi-dimensional VN embedding model and proposes a topology-aware method to embed multi-dimensional VNs.It transforms containers and simulation networks,and after the process of coarsening,partitioning,and uncoarsening multi-dimensional VN topology,a topology-aware repeatable embedding scheme is used to complete the embedding stage.Small-,1,000-and 10,000-scale VNE simulation experimental results demonstrate that our method outperforms eight comparison approaches.Remarkably,this method improves acceptance ratio,revenue,and revenue-cost ratio by 12.19%,12.19%,and 57.45% on average,respectively,and reduces average cost rate by 10.61%.Furthermore,real-world Open Stack-based embedding experimental results reveal the ability of our method to efficiently reduce communication costs by up to 45.93% and 63.43% for download and upload,respectively.Then,after embedding VNs in Cyber Range,this thesis studies virtual link emulation for multi-dimensional VN to build a high-fidelity link emulation environment in Cyber Range.Aiming at the problem that traditional methods possess poor virtual link characteristics emulation accuracy,this thesis proposes a virtual link emulation algorithm using Kalman prediction theory.It considers inherent network traffic errors of hardware and software in the physical platform,thereby improving link emulation fidelity.Subsequently,this thesis builds a virtual link emulation system based on multiple virtualization technologies.This system utilizes our virtual link emulation algorithm,integrates deployment,link-emulation,and emulator modules,and implements a measurement scheme for packet delay and packet loss rate to provide calculation data for our virtual link emulation algorithm.Open Stack-based evaluation results show that our emulation system can flexibly construct VN emulating a Bei Dou based satellite network scenario,and set the expected bandwidth for each virtual link within 5% errors.In addition,our virtual link emulation algorithm outperforms the traditional method by 11.48% and 49.51% on average in terms of packet delay and packet loss rate,respectively.Next,after the VN construction,this thesis studies multiple virtual resource management in each host machine where the VN is located to improve the execution efficiency of emulation tasks in Cyber Range.Aiming at the inefficiency of traditional methods in collaborative management of multiple resources,this thesis proposes a dynamic resource management algorithm based on deep deterministic policy gradients.It considers multiple resources,including CPU,memory,and bandwidth,and combines the proposed action refinement algorithm to achieve dynamic resource configuration for virtual nodes.Subsequently,this thesis builds an adaptive resource management system that integrates monitoring,calculation,and execution modules to dynamically schedule CPU,memory,and bandwidth in hardware and lightweight virtualization technologies.Open Stack-based lightweight and heavyweight experimental results show that our scheme effectively utilizes physical resources and considerably improves the performance of benchmarks in virtual machines and containers.Compared with the completion time results of the two comparison methods,our scheme can reduce by 41.22% and 38.45% on average in lightweight and heavyweight evaluations.Last,after the multiple virtual resource management,this thesis studies the multiobjective migration method in the cluster where the VN is located to improve multiobjective resource efficiency for Cyber Range.Aiming at the lack of careful consideration in optimization objectives of traditional migration methods,this thesis establishes a multiobjective VN migration model regarding energy,communication,migration,and service level agreement violation(SLAV).It proposes a dynamic migration algorithm based on double deep Q-learning.It uses deep neural networks to deal with ample state space to schedule migrations dynamically and adopts an action selection algorithm to reduce action space for the learning model to promote the algorithm’s efficiency in a large-scale network environment.Small-and large-scale VN migration experimental results demonstrate that our method is much more efficient than the two comparison methods.Remarkably,our method decreases average SLAV and communication cost by 23.75% and 9.40%,respectively,and reduces total cost by 15.61% on average.Furthermore,Open Stackbased VN migration experimental results reveal the ability of our method to make full use of physical computation and network resources and reduce the completion time of computation-and network-intensive benchmarks by 11.35% and 10.31%,respectively.