Scalable Analytics for IaaS Cloud Availability
Scalable Analytics for IaaS Cloud Availability
复制标题
DOI：
10.1109/tcc.2014.2310737
复制
发表时间：
2014-01
影响因子：
6.5
通讯作者：
R. Ghosh;F. Longo;Flavio Frattini;S. Russo;Kishor S. Trivedi
中科院分区：
文献类型：
--
作者：
R. Ghosh;F. Longo;Flavio Frattini;S. Russo;Kishor S. Trivedi
在大型基础设施即服务(IaaS)云中，组件故障非常常见。此类故障可能会导致偶尔的系统停机，并最终违反云服务可用性的服务级别协议(SLA)。底层基础设施的可用性分析对于服务提供商设计能够提供定义的SLA的系统以及评估现有SLA的功能非常有用。本文提出了一种可扩展的随机模型驱动的方法来量化大规模IaaS云的可用性，其中故障通常通过在三个池之间迁移物理机来处理：热(运行)、热(打开，但未准备好)和冷(关闭)。由于整体模型不适用于大系统，我们使用基于交互马尔可夫链的方法来证明分析的复杂性和求解时间的降低。这三个池由互动子模型建模。利用不动点迭代法解决了它们之间的依赖关系，并证明了其解的存在性。将所提出的方法与整体模型得到的解析-数值解进行了比较。我们表明，相互作用子模型引入的误差是微不足道的，并且我们的方法可以处理非常大的IaaS云。对所提出的模型也考虑了模拟解，并对各种方法的求解时间进行了比较。
In a large Infrastructure-as-a-Service (IaaS) cloud, component failures are quite common. Such failures may lead to occasional system downtime and eventual violation of Service Level Agreements (SLAs) on the cloud service availability. The availability analysis of the underlying infrastructure is useful to the service provider to design a system capable of providing a defined SLA, as well as to evaluate the capabilities of an existing one. This paper presents a scalable, stochastic model-driven approach to quantify the availability of a large-scale IaaS cloud, where failures are typically dealt with through migration of physical machines among three pools: hot (running), warm (turned on, but not ready), and cold (turned off). Since monolithic models do not scale for large systems, we use an interacting Markov chain based approach to demonstrate the reduction in the complexity of analysis and the solution time. The three pools are modeled by interacting sub-models. Dependencies among them are resolved using fixed-point iteration, for which existence of a solution is proved. The analytic-numeric solutions obtained from the proposed approach and from the monolithic model are compared. We show that the errors introduced by interacting sub-models are insignificant and that our approach can handle very large size IaaS clouds. The simulative solution is also considered for the proposed model, and solution time of the methods are compared.