Modeling And Optimization For Software-Defined Multimedia Service

With the significant advances of network technology and the fast improvement of network service,network traffic has grown explosively in recent years.The construc-tion speed of network limited by Moore's law can't match the development of different kinds of network service.Among all the traffic,multimedia traffic is undergoing a rapid growth as the biggest one.According to the research by Cisco,it will be 80%of all network traffic in 2020.Blind expansion of the current network is unable to meet the increasing demand of users,the intelligent network which could provide the perception of users' behavior and network information can use the limited network re-sources efficiently to improve the Quality of Service(QoS).However,the forwarding plane of traditional IP network is transparent for the application layer,which causes the lack of capacity of network perception so that it is hard to deploy customized algo-rithms for different service in a shared network.Software-Defined Networking(SDN)has been fast emerging as a promising network technology for building next-generation services and networks.The unique architecture which separates the control plane from data plane could have an overall view and improve the cooperation of network layer and application layer.Meanwhile,the centralized control plane can also provide an open platform for intelligent algorithms deploying.Therefore,this thesis focuses the research on modeling and optimization for multimedia service over SDN architecture based on Markov decision process and approximate dynamic programming.With the aid of neuro-dynamic programming and deep learning method,the proposed optimiza-tion problems have been effectively solved.The main contributions of this thesis are summarized as follows:1)Researching on joint admission control and routing via approximate dynamic pro-gramming for streaming video over SDN.In a bandwidth limited specialized net-work for video service,video server provides streaming video transmission for multi-users with rewards,the keys of this optimization problem are reasonable admission control and efficient routing algorithm.This thesis employs the tech-nique of network virtualization to build a virtual SDN network for streaming video service,in order to deploy the optimization algorithm flexibly.In theory,the pro-posed optimization problem is formulated as a Markov Decision Programming(MDP)for maximizing the overall rewards.In consideration of the curses of di-mensionality caused by the large state space,the method of approximate dynamic programming(ADP)and kernel-based function approximation is invoked to ob-tain an approximate optimal solution.Based on FlowVisor,POX controller and Mininet,an emulation platform is constructed to verify the effectiveness of pro-posed ADP based algorithm.The experimental results show the performance im-provement of the proposed algorithm by comparing it with Q-learning algorithm and Open Shortest Path First(OSPF)based algorithm.2)Constructing an open architecture for DASH video routing and bitrate adaption in SDN by neuro-dynamic programming.Different with the traditional video coding,DASH(Dynamic Adaptive Streaming over HTTP)video has the charac-teristic of multi segments and multi bitrate copies,which bring big optimization feasibility.The traditional optimizations for DASH video transmission concen-trate on the terminals and servers.With the aid of feedback information from user terminals,DASH video server could provide dynamic transmission in dynamic bitrates to match users' behavior.Under this optimization mechanism,the net-work information is still unused and potential delay of the feedback information may also reduce the performance.The flexible architecture of SDN brings the fea-sibility of DASH video transmission optimization in network layer.In this thesis,the optimization of DASH video routing and bitrate adaption are both considered in SDN,the proposed optimization problem is formulated as a MDP problem.In order to solve the curses of dimensionality,neuro-dynamic programming(NDP)method is employed to conceive an online learning framework for DASH video service.According to the emulation results,the proposed algorithm has more excellent performance compared with OSPF based algorithm.3)Studies on layer adaption and routing for scalable video transmission in SDN us-ing deep learning based NDP.As a burgeoning video coding standard,the studies on scalable video coding(SVC)transmission optimization have prominent signif-icance.By encoding the video source to one base layer and serval enhancement layers,scalable video could be transmitted to uses with different layers based on user requirement.Limited by traditional network architecture,most the ex-isting SVC transmission optimization are across-layer collaboration,and within the framework of SDN,more researchers focus on the single-user optimization and multi-cast service.This thesis studies the SVC transmission optimization for multi-users in bandwidth limited SDN.With the real time network information obtained by SDN controller,each user's request of scalable video transmission is intelligently handled to bring the maximum utilization of network resource.This optimization problem is formulated as a Markov decision process.In order to solve the curses of dimensionality and remove the instability of manual feature extraction,the deep learning based NDP method is proposed to obtain an approx-imate optimal solution.The emulation results show that the proposed algorithm performs much better than the basic NDP based algorithm.For different video formats of the above three optimization problems,SDN ar-chitecture is employed to build an open system model for multimedia service and the transmission problems are formulated as Markov decision process.With the aid of serval optimization algorithms,the problems of admission control,routing and bitrate adaption are jointly solved in proposed system model and theory model.These three optimization problems are progressive in objectives and methods.The diversity of ob-jects and methods make the modeling and optimization for multimedia in this thesis have generality,which verifies the effectiveness and openness of the proposed multi-media optimization platform based on SDN and MDP.