| Recently,mobile services are rapidly developed due to the improvement of people's living standards.A large number of new applications(e.g.,vehicular network services,augmented reality,and smart cities)emerge,which provides extreme convenience to our lives while putting higher requirements on mobile networks.On the one hand,most of these applications have high demands on physical resources.The gaps between resourceshungry applications and resources-constraint mobile terminals are prominent.On the other hand,emerging applications with diverse needs result in rapid traffic growth.Dramatic performance enhancement needs to be done to cope with the particular needs in different scenarios.To address the above-mentioned issues,the industrial and academic communities develop the 5G network,in which the virtualization and cloud computing are the key techniques.Virtualization focuses on building an open and flexible network platform that enables us to customize the network,improving resource utilization,and reducing the operational cost.The cloud computing paradigm enhances the capacity of user terminals by computation offloading.The battery life is also extended since the energy consumption is reduced.This dissertation studies the deployment of the virtualized mobile network.We mainly focus on the following topics:(1)how to design an efficient and topologyindependent task offloading mechanism in a multi-tier MEC network;(2)how to deploy the MEC nodes in a virtualized MEC network to improve the network efficiency and reduce the network cost;(3)how to deploy a service chain to serve the time-and spacevarying user demands.The main contributions are list as follows:First,the offloading strategy and resource allocation mechanism in a multi-tier network: To improve the availability of current offloading algorithms,we study the task offloading in a generalized network topology.Offloading decision,transmitted power control,and computation&communication load distribution are jointly optimized to reduce energy consumption and latency.A distributed offloading scheme based on the game theory is proposed to overcome the issue of the curse of dimensionality.We use a marginal payoff function to ensure convergence.The optimality of the proposed algorithm is proved by the potential function.Furthermore,to adapt to the time-varying environment,a fast-converged offloading algorithm is designed by introducing an approximate factor.We theoretically demonstrate the impacts brought by the approximate factor on both the accuracy and the time complexity.The simulations show that the proposed algorithms can efficiently improve network efficiency compared to baseline algorithms.Moreover,the proposed fast-converged algorithms can reduce 72.1% of iterations with 50 users when the approximate factor is 1.2.The reduction enlarges to 84.4% when the approximate factor increases to 1.6.Second,the MEC node deployment and resource allocation in the virtualized MEC network: To decrease the operational cost and service latency,the MEC node placement,along with the communication and computational resource allocation,are commonly considered during the optimization.However,by using the dynamic sleep scheduling technique,the network cost can be further reduced by scheduling idle nodes and links to sleep.However,the integration of dynamic sleep scheduling significantly increases the problem scale.Moreover,the network connectivity may be lost during the network reconfiguration,which may make the traditional centralized methods unsuitable for real implementation.To cope with the above-mentioned issues,this dissertation combines the cost-sharing game and the self-routing game to develop a mixed game.Then,based on the mixed game,a distributed cost-latency aware algorithm is proposed.We prove the mixed game is conditionally an ordinary potential game.The close-forms of the convergence speed and the efficiency are provided.To further improve the accuracy,an improved algorithm based on the Stackelberg routing is developed.We theoretical analyze the algorithm performance under special cases(i.e.,linear latency model and parallel routing).Finally,the efficiency of the proposed algorithms is validated through simulations.Compared to conventional MIP(mixed integer programming)-based solver(e.g.,genetic algorithm),the proposed algorithms take more advantages in delayinsensitive scenarios.Third,the dynamic deployment of service function chains using the deep reinforcement learning algorithm: To improve the efficiency of existing learning-based algorithms,we propose a hybrid learning framework for service function chaining,in which the placement of VNF instances and the routing of user flows are decoupled during the training.The learning agent only focuses on the optimization of VNF instance placement,which shrinks the dimension of the action space,and thus improves the learning efficiency.During the training,the flow routing is conducted by another module.The results of routing would be fed back to the learning agent as the reward signal to ensure the joint optimization.To emphasize the low complexity of flow routing,a distributed scheme based on the game theory is designed,in which an additional player is introduced to cope with the high complexity brought by the global constraint.The performance on both the convergence and the accuracy of the proposed algorithm is also illustrated.Through simulations,we show the superiority of the proposed algorithm,which can outperform baselines by 9%-15% on the overall system cost.This dissertation focusses on the task offloading,the MEC node deployment,and the service function chaining,which are key issues for the deployment of virtualized mobile networks.We are aimed at reducing the network cost and the service latency by scheduling the network resources accordingly.Meanwhile,we emphasize the scalability,the time complexity,and performance traceability of the proposed algorithms,which are important during the real-world deployment.This research can serve as a reference for the deployment of future mobile networks. |
PDF Full Text Request