【作者】 刘宇轩；

【导师】 张三峰； 翟明玉；

【作者基本信息】 东南大学 ， 计算机技术（专业学位）， 2021， 硕士

【摘要】 传统上基于数据中心的云计算模式在服务日益增加的移动边缘终端用户时会产生大量的骨干网络流量,并导致难以忍受的响应延时。移动边缘计算(Mobile Edge Computing,MEC)通过在基站之类的设备附近部署小规模的服务器集群,将移动应用的计算任务从数据中心迁移到网络边缘,降低了骨干网络的负载也提高了服务的实时性。在MEC网络中使用网络功能虚拟化(Network Function Virtualization,NFV)技术为移动应用提供网络服务可以有效保证数据传输的安全性并提升服务质量。但与数据中心网络不同,在MEC网络中基于NFV提供网络服务面临额外的挑战:MEC网络中的服务器集群是地理上分布的,仅有有限的计算资源,且移动用户具有严格的服务质量要求和很强的移动性,这对虚拟网络功能的布置提出了更严苛的要求。为此,本文针对MEC网络中基于NFV的多播,提出了一种启发式的NFV服务功能链嵌入和资源分配算法以及一种基于用户移动预测的多播网络动态调整方法,通过优化虚拟网络功能放置和资源分配方案改善网络服务质量。本文的主要工作包括:(1)形式化地定义了MEC网络中基于NFV的多播网络构建问题和多播网络动态调整问题,使用整数线性规划对其建模,并证明其为NP-hard问题。(2)针对MEC网络中基于NFV的多播网络构建问题,提出了一个两阶段的启发式算法,用于确定虚拟网络功能的部署位置和多播流流量的路由路径。算法的第一阶段在给定延时约束下,寻找最小成本的服务链嵌入方案。第二阶段以最小化延时差异为目的,确定服务功能链中最后一个虚拟网络功能的部署位置以及该位置与目的地之间的路由路径。证明了该算法可以在更优的时间复杂度内找到构建问题的可行解,并通过实验对比分析,验证了算法在不同网络拓扑上的性能。(3)针对MEC网络的动态性问题,提出了一个基于用户移动预测的多播网络动态调整方法。该方法使用基于常微分循环神经网络的移动预测模型,根据用户历史移动轨迹预测用户的下一个位置,并根据预测结果对已部署的服务功能链和路由流量的多播树进行动态调整。通过实验验证了不同的模型在移动预测上的性能,并根据实验结果分析了动态服务调整带来的用户体验改善。验证结果显示本文所提移动预测模型相比对比模型提升了4.91%～11.40%的准确率和56m～264m的绝对距离误差;本文所提的动态调整策略相比不调整多播网络的策略,在不同用户数量下可以降低41.03%～87.5%的服务质量违反概率,相比不缩减备份成本的调整策略减少6.83%～9.66%的备份成本。总之,本文提出的启发式算法,可以在有限的时间内,确定NFV服务功能链的嵌入策略和多播流量路由路径,在保证服务质量的同时最小化多播网络的构建成本。将常微分循环神经网络应用于移动预测,可以有效的预测未来某个时刻用户所处的位置。基于移动预测的动态调整算法,可以通过实时调整基于NFV的多播网络,改善用户移动后的用户体验。论文成果可以有效改善MEC网络中基于NFV的多播服务质量,同时减少基于NFV的多播网络构建成本。更多还原

【Abstract】 When serving increasing mobile users,the traditional cloud computing paradigm based on the data center will lead to backbone network congestion and cause high response delay which is unbearable for some real-time applications.Mobile Edge Computing(MEC)migrates the computing tasks of mobile applications from the data center to the edge of network by deploying small-scale server clusters near base stations that are within the proximity of mobile user,reducing response delay for real-time applications.Providing network services for mobile applications using Network Function Virtualization(NFV)technology in MEC can ensure the security of data transmission and improve the quality of service.However,unlike data center,providing network services using NFV in MEC faces additional challenges: the server clusters in MEC are geographically distributed,with limited computing resources,and mobile users have strict service quality requirements and strong mobility,which puts forward more stringent requirements on the placement of virtual network functions.To this end,considering providing network services for NFV-enabled multicasting in MEC networks,a heuristic algorithm for service function chain embedding and an algorithm based on user mobility prediction for dynamic adjustment are proposed,which improves network service quality by optimizing virtual network function placement and resource allocation.The main work includes:(1)The NFV-enabled multicasting network construction problem and the NFV-enabled multicasting network dynamic adjustment problem are defined formally,which is modeled using Integer Linear Programming.It is proved that the above two problems are NP-hard problem.(2)A two-stage heuristic algorithm for the problem of NFV-enabled multicasting network construction is proposed,which is used to determine the placement of virtual network function and the routing path of multicasting flow.The first stage of the algorithm finds the embedding scheme of the service function chain under the given delay constraint,minimizing the construction cost.The second stage determines the placement of the last network function in the service function chain and the routing path between that location and all destinations,minimizing the delay variations.It is proved that the algorithm can find a feasible solution to the construction problem within a better time complexity.The performance of the proposed algorithm on different network topologies is evaluated through experiments.(3)A heuristic algorithm based on user mobility prediction for the dynamic adjustment of the constructed multicasting network is proposed.This algorithm uses an Ordinary Differential Equation Recurrent Neural Network to predict the user’s next position based on the user’s historical trajectories,and dynamically adjust the deployed service function chain and multicasting tree using the prediction result.The performance of the different mobility prediction models and the user experience improvement brought by dynamic adjustment are evaluated through experiments.The results show that the mobile prediction model proposed improves the accuracy of 4.91%～11.40% and the absolute distance error of 56m～264m compared with other models.The dynamic adjustment strategy proposed can reduce the probability of service quality violation by 41.03%～87.5% compared with no adjustment strategy,and reduce the backup cost by 6.83%～9.66% compared with the adjustment which do not reduce backup cost.In short,the heuristic algorithm proposed can determine the embedding strategy of the NFV service function chain and the routing path within a limited time,minimizing the construction cost of the multicasting network while ensuring the quality of service.The proposed user mobility prediction model based on the Ordinary Differential Recurrent Neural Network can effectively predict the location of the user at a certain moment.The dynamic adjustment algorithm based on user mobility prediction can proactively reconstruct the NFVbased multicasting network to improve the user experience after the user moves.The contributions can effectively improve the quality of NFV-based multicasting in the MEC network,while reducing the construction cost of NFV-based multicasting network.更多还原