Abstract: Focusing on the admission control problems of media delivery network in Software Defined Network (SDN), an admission control scheme of comprehensively optimizing the service performance of service nodes and transmission links was proposed. The proposed scheme used SDN controller's abilities of directly controlling routers and perception of entire network, to jointly optimize the service performance of the service nodes on the application layer and the transmission links between the service nodes and the users on the network layer, and the influences of link congestion to data transmission and quality of service were reduced. Firstly, the admission control process of the SDN service system was modeled as a Partially Observable Markov Decision Process (POMDP). Secondly, the observation-based randomized policy was used as the admission control policy of the system. Finally, the policy-gradient algorithm was used to optimize the admission control policy, and the optimal policy of the model was obtained. The simulation results show that, compared with the best effort service policy, the POMDP-based optimal admission control policy improves system performance by 10%, which illustrates the effectiveness of the proposed approach.

Key words: Software Defined Network (SDN), admission control, Partially Observable Markov Decision Process (POMDP), randomized policy, policy-gradient approach

摘要： 针对软件定义网络（SDN）环境下的媒体分发网络的接入控制问题，提出了一种综合考虑服务节点和传输链路服务性能优化问题的接入控制方案。该方案利用SDN控制器对路由器的直接管控和对全网的感知能力，对应用层上服务节点服务性能和网络层上服务节点到用户之间的传输链路的服务性能进行联合优化，减少了链路拥塞对数据传输和用户服务质量的影响。首先，为SDN服务系统的接入控制过程建立部分可观Markov决策过程（POMDP）模型；然后，使用基于观测的随机策略作为系统的接入控制策略；最后，通过策略梯度算法对接入控制策略进行优化，求解出模型的最优策略。仿真结果表明，与尽力而为的服务策略相比，基于POMDP模型的最优接入控制策略使系统性能提高了10%，验证了所提方法的有效性。

关键词: 软件定义网络, 接入控制, 部分可观Markov决策过程, 随机策略, 梯度优化