基于软件定义网络的集群虚拟云资源分配

发布时间：2018-05-04 23:02

本文选题：云计算 + 网络性能　；参考：《西南大学》2017年硕士论文

【摘要】：随着云计算平台越来越庞大并且呈现动态化,有效地管理整个云计算数据中心的资源同时满足用户应用的资源需求成为云提供商和租户共同的诉求。云计算的目标是,租户可以根据随收随付模式租赁计算资源,云提供商可以有效地管理计算资源以保证服务的可扩展性和有效性。网络虚拟化技术的出现为云计算资源的高效管理带来了契机,通过“抽象”整个云数据中心的物理基础设施形成巨大的虚拟“资源池”,满足多用户同时共享资源的需求,大大降低云资源管理难度和维护成本。代表未来网络发展方向的软件定义网络拥有整个网络的全局视图,可以细粒度地操纵整个云数据中心资源,使云计算资源分配任务变得更加有效。然而,由于租户的大量服务和应用程序通常共享同一资源池,使得数据中心环境变得非常复杂,因此,满足多个用户的计算请求而进行有效的资源分配,同时保障整个网络的性能将成为巨大的挑战。为了解决这个问题,首先,本文构建一种基于软件定义网络的fat-tree数据中心虚拟资源分配网络模型。其次,本文提出一种集群虚拟网络(Clustered Virtual Network,CVN)策略和动态heapsort算法寻找含有最合适数量的虚拟机(VM)的交付单元(pod),以有效地利用单个pod中的资源构建集群虚拟网络。然而,当单个pod找不到足够的资源时,将联合多个pod构建跨pod集群虚拟网络(Crossing Pod-CVN,CP-CVN),并提出一种max-min策略以进行跨pod的虚拟机分配。进而,在虚拟机分配的基础上,本文进行网络资源分配,并将此网络资源分配优化问题公式化为线性规划问题(Linear Programming,LP)以最大化网络吞吐量,并提出一种近似算法解决此线性规划问题,保障网络的整体性能。理论分析表明本文所提出的近似算法可以有效地解决软件定义网络中的集群虚拟云资源分配问题并保证网络性能。最后,本文通过网络仿真平台Mininet部署了仿真实验进行方法评估。Mininet是一个被学术界广泛肯定和采用的网络仿真器,它使用轻量级虚拟化在单个系统(如计算机或虚拟机)上部署大型网络。本文从拓扑结构、通信模式、资源分配模式、数据流生成等多维度设计实验,以验证所提出的算法(统称为CVN算法)的有效性。实验结果表明,本文所提出的CVN算法较Cloud Mirror、Oktopus和MILPFlow算法在网络吞吐量和平均传输时延方面有更好的性能表现,证实理论分析的正确性。
[Abstract]:As the cloud computing platform becomes more and more large and dynamic, it becomes a common demand of cloud providers and tenants to manage the resources of the whole cloud computing data center effectively and meet the resource needs of users. The goal of cloud computing is that tenants can rent computing resources according to pay-as-you-go model, and cloud providers can effectively manage computing resources to ensure the scalability and effectiveness of services. The emergence of network virtualization technology brings opportunities for the efficient management of cloud computing resources. By "abstracting" the physical infrastructure of the entire cloud data center, a huge virtual "resource pool" is formed to meet the needs of multi-users sharing resources at the same time. Greatly reduces the cloud resources management difficulty and the maintenance cost. The software definition network, which represents the future direction of network development, has the global view of the whole network, which can manipulate the whole cloud data center resource fine-grained and make the resource allocation of cloud computing more effective. However, because a large number of services and applications of tenants often share the same resource pool, the data center environment becomes very complex, so efficient resource allocation is made to satisfy the computing requests of multiple users. At the same time, ensuring the performance of the entire network will become a huge challenge. In order to solve this problem, firstly, this paper constructs a virtual resource allocation network model of fat-tree data center based on software defined network. Secondly, this paper proposes a cluster virtual network (cluster Virtual network CVN) strategy and dynamic heapsort algorithm to find delivery units containing the most appropriate number of virtual machines (VMs), so as to effectively utilize the resources in a single pod to construct a cluster virtual network. However, when a single pod can not find enough resources, it will combine multiple pod to construct Crossing Pod-CVN CP-CVN, and propose a max-min strategy to allocate virtual machines across pod. Furthermore, on the basis of virtual machine allocation, the network resource allocation problem is formulated as linear programming problem (LPP) to maximize network throughput. An approximate algorithm is proposed to solve the linear programming problem to ensure the overall performance of the network. Theoretical analysis shows that the proposed approximate algorithm can effectively solve the cluster virtual cloud resource allocation problem in the software defined network and ensure the network performance. Finally, this paper deploys the simulation experiment through the network simulation platform Mininet to evaluate. Mininet is a network simulator that is widely recognized and adopted by the academic circles. It deploys large networks on a single system, such as a computer or virtual machine, using lightweight virtualization. In order to verify the effectiveness of the proposed algorithm (collectively known as CVN algorithm), this paper designs multi-dimensional design experiments from topology, communication mode, resource allocation mode, data stream generation and so on. The experimental results show that the proposed CVN algorithm has better performance in network throughput and average transmission delay than the Cloud Mirror-Oktopus and MILPFlow algorithms. The correctness of the theoretical analysis is verified.
【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP308;TP393.0

【参考文献】