【作者】 杨旭；

【导师】 武穆清；

【作者基本信息】 北京邮电大学 ， 信息与通信工程， 2021， 硕士

【摘要】 当今时代,互联网和云计算的蓬勃发展让数据中心扮演着非常重要的角色。数据中心网络因为可靠性和健壮性的要求,服务器间拥有多条路径,但是传统的数据中心网络拥塞控制机制无法充分利用这些可选路径,也无法有效地解决数据中心的拥塞问题。所以软件定义网络(Software Defined Networks,SDN)网络架构逐渐在数据中心网络中被广泛地使用。SDN可以获取网络的全局信息,更好地监测网络状况并且制定拥塞控制策略。但是SDN网络架构也有其内在缺点,如果没有合理的流量调度策略或者控制器负载过重,还是会导致网络拥塞。所以本论文从流量调度和控制器拥塞减免两方面入手来解决软件定义数据中心网络的拥塞问题。(1)提出基于分段标签下发的控制器拥塞控制策略。传统的OpenFlow协议会在流量路径上的每个交换机上都去下发安装流表项,这对控制器造成了很大的负担,增大了控制器的拥塞,所以本文使用在多节点下发MPLS标签的方式来减少控制器流表的下发。本文推导出了一个数学模型用来描述最终的优化目标,这是一个非确定性多项式难题(NP-hard问题),本文使用贪心算法来取得优化目标函数的局部最优解,选择合适的交换机下发带MPLS标签栈的流表信息。最终实验的结果和传统的OpenFlow下发流表的方式比较可以减少60%流表项的下发,并且在大流量的环境下可以使控制器的延时减少25%左右,实现了通过降低控制器的负载和控制器处理时延的方式来减轻网络整体的拥塞。(2)提出基于模糊逻辑推理的流量调度策略。使用模糊逻辑推理模型,将候选路径的链路利用率和路径跳数作为输入值,经过模糊化、规则推理、解模糊化的步骤,推理出该条路径的评分值,最后取所有路径的评分最大值,则该条路径为最终所选的路径。仿真实验结果表明,模糊逻辑算法在吞吐量、负载均衡和平均往返时延方面比ECMP、Hedera等算法有显著的改进,能够有效降低网络拥塞。更多还原

【Abstract】 Nowadays,with the rapid development of Internet and cloud computing,data center plays a key role.Due to the reliability and robustness of data center network,there are multiple paths between servers.However,the traditional congestion control mechanism of data center network can not make full use of these alternative paths and effectively solve the congestion problem of data center.Software defined networks(SDN),a new network architecture,has been widely used in data center network.According to the global information obtained by SDN,we can better monitor the network condition and formulate congestion control strategy.But SDN network architecture also has its inherent shortcomings.If there is no reasonable traffic scheduling strategy or the controller is overloaded,it will still lead to network congestion.Therefore,this paper from the two aspects of traffic scheduling and controller congestion relief to solve the congestion problem of software defined data center network.(1)A congestion control strategy based on multi node label distribution is proposed.The traditional OpenFlow protocol will distribute and install flow table entries on each switch on the traffic path,which causes a great burden on the controller and increases the congestion of the controller.In this paper,we use the method of MPLS label distribution in multiple nodes to reduce the distribution of controller flow table,and reduce the load of the controller.In this paper,a mathematical model is derived to describe the final optimization objective,which is a non-deterministic polynomial problem(NP hard problem),so the greedy algorithm is used to obtain the local optimal solution of the optimization objective function,and the flow table information with MPLS label stack is distributed by the appropriate switch.Finally,the experimental results show that compared with the traditional OpenFlow method,it can reduce the distribution of flow table entries by 60%,and reduce the delay of the controller by about 25%in the environment of large traffic.It can reduce the overall network congestion by reducing the load of the controller and the processing delay of the controller.(2)A traffic scheduling strategy based on fuzzy logic reasoning is proposed.Using the fuzzy logic reasoning model,the link utilization and hop number of the candidate path are taken as the input values.After the steps of fuzzification,rule reasoning and defuzzification,the score value of the path is deduced.Finally,the maximum score of all paths is taken,and the path is our final selected path.The simulation results show that the fuzzy logic algorithm is better than ECMP and Hedera in throughput,load balancing and average round-trip delay,which can effectively reduce network congestion.更多还原