数据中心网络结构及其调度优化算法研究

发布时间：2018-04-05 05:13

本文选题：数据中心　切入点：网络结构HRN　出处：《浙江理工大学》2017年硕士论文

【摘要】：为了适应新兴应用模式的发展和需求,数据中心网络正发生着深刻的变革,不仅表现在规模扩展性的提升和成本控制上,还体现在资源调度策略等方面。网络拓扑结构和资源调度算法是现代数据中心的研究重点,现有的网络结构虽然具有较好的扩展性和容错性,但存在扩展规模受限于服务器网络端口数的问题,不能同时兼顾性能和经济成本。其次,随着数据中心规模和用户数量的急剧增大,当前资源调度算法在面对用户多类别请求时不能有效的平衡执行效率和负载均衡。针对上述问题,本文具体研究以下内容:(1)针对当前数据中心网络扩展性受限于服务器网络端口数的问题,本文提出一种新型高扩展低端口的数据中心网络结构HRN(Hyper Ring Network)。HRN基于低端口普通服务器和交换机,以分层递归定义的形式构建大规模数据中心网络结构。首先定义HRN结构的编码规则和构建规则,根据其构建方式得出其拓扑属性并证明;然后基于HRN结构设计最短路径路由、并行路径路由和容错路由,来保证数据中心的通信性能;最后实验模拟HRN拓扑及其路由算法,并从拓扑性和可靠性两方面与其它结构进行对比。实验结果表明该结构能以较低的服务器和交换机比例来实现大规模网络拓扑,具有高度扩展性,且有效降低了构建成本,同时高效路由算法使结构能提供良好的吞吐量和强可靠性。(2)数据中心作为管理和调度资源的共享平台,必须具备高效的资源调度策略。针对调度过程中系统的负载失衡问题,本文结合上述结构HRN提出一种基于Max-Min算法与蚁群算法融合的数据中心资源调度优化算法:(1)基于数据中心的资源调度模型设计调度优化目标,然后综合考虑用户请求的时间约束、可靠性、通信带宽等指标,制定资源约束函数;(2)针对传统蚁群算法在解决调度问题时存在的不足,本文改进了状态转移概率公式、局部信息素、全局信息素更新公式,然后就资源负载不均问题提出了负载调整因子,并将其加入信息素更新公式,从而保证调度过程的负载均衡;(3)就蚁群算法前期收敛慢的问题,提出将Max-Min算法与改进后的蚁群算法相结合,因为Max-Min算法具有处理效率高且负载均衡效果好的优点,所以算法前期利用Max-Min进行全局寻优,用得到的最优解来初始化蚁群算法的信息素分布,从而加快算法收敛速度;(4)用CloudSim模拟本文算法,并与Max-Min和蚁群算法进行对比,实验结果表明本文算法加快了整体的调度速度,能在保证系统负载均衡的同时以较短的时间完成调度。
[Abstract]:In order to adapt to the development and demand of the emerging application mode, the data center network is undergoing profound changes, not only in the scale expansion and cost control, but also in the resource scheduling strategy.Network topology and resource scheduling algorithms are the focus of modern data center research. Although the existing network structure has good scalability and fault-tolerance, there is a problem that the expansion scale is limited by the number of server network ports.Performance and economic costs cannot be considered at the same time.Secondly, with the rapid increase of data center size and the number of users, the current resource scheduling algorithm can not effectively balance execution efficiency and load balance in the face of multi-class user requests.In view of the above problems, this paper specifically studies the following contents: 1) aiming at the problem that the expansibility of the current data center network is limited by the number of server network ports,In this paper, a new type of data center network structure, HRN(Hyper Ring Network).HRN, with high expansion and low ports, is proposed, which is based on low port common server and switch, and constructs large-scale data center network structure in the form of hierarchical recursive definition.Firstly, the coding rules and construction rules of HRN structure are defined, and its topological properties are obtained and proved according to its construction mode. Secondly, the shortest path routing, parallel path routing and fault-tolerant routing based on HRN structure are designed to ensure the communication performance of the data center.Finally, the HRN topology and its routing algorithm are simulated and compared with other structures in terms of topology and reliability.The experimental results show that the structure can realize the large-scale network topology with a low ratio of servers and switches, and has a high scalability, and effectively reduces the construction cost.At the same time, the efficient routing algorithm enables the structure to provide good throughput and strong reliability. The data center is a shared platform for resource management and scheduling, and must have an efficient resource scheduling strategy.Aiming at the problem of system load imbalance during scheduling,In this paper, a data center resource scheduling optimization algorithm based on the combination of Max-Min algorithm and ant colony algorithm is proposed, which is based on the above structure HRN. The resource scheduling model based on the data center is used to design the scheduling optimization objectives, and then the time constraints of user requests are considered synthetically.Aiming at the shortcomings of traditional ant colony algorithm in solving scheduling problems, this paper improves the formula of state transition probability, local pheromone and global pheromone updating.Then, the load adjustment factor is put forward for the problem of uneven load of resources, and the pheromone updating formula is added to ensure the load balance of scheduling process (3) the problem of slow convergence in the early stage of ant colony algorithm.This paper proposes to combine the Max-Min algorithm with the improved ant colony algorithm. Because the Max-Min algorithm has the advantages of high processing efficiency and good load balancing effect, Max-Min is used for global optimization in the early stage of the algorithm.The optimal solution is used to initialize the pheromone distribution of the ant colony algorithm, so as to speed up the convergence of the algorithm. The algorithm is simulated by CloudSim and compared with Max-Min and ant colony algorithm. The experimental results show that the proposed algorithm accelerates the overall scheduling speed.At the same time, the scheduling can be completed in a short time while ensuring the load balance of the system.
【学位授予单位】：浙江理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP308

【参考文献】