Abstract
Cloud based systems are inherently large scale. Failures in such a large distributed environment are quite common phenomena. To reduce the overall Cloud downtime and to provide a seamless service, providers need to assess the availability characteristics of their data centers. Such assessments can be done through controlled experimentations, large scale simulations and via analytic models. In the scale of Cloud, conducting repetitive experimentations or simulations might be costly and time consuming. Analytic models, on the other hand, can be used as a complement to small scale measurements and simulations since the analytic results can be obtained quickly. However, accurate analytic modeling requires dealing with large number of system states, leading to state-space explosion problem. To reduce the complexity of analysis, novel analytic methods are required. This chapter introduces the reader to a novel approach using interacting analytic sub-models and shows how such approach can deal with large scale Cloud availability analysis. The chapter puts the work in perspective of other existing and ongoing research in this area, describe how such approach can be useful to Cloud providers, especially in the case of federated scenarios, and summarize the open research questions that are yet to be solved.
TopIntroduction
Cloud computing is a model of Internet-based computing. An IaaS Cloud, such as Amazon EC2 and IBM SmatCloud Enterprise™ (Amazon EC2: www.ibm.com/services/us/en/cloud-enterprise/, 2011) delivers, on-demand, operating system (OS) instances provisioning computational resources in the form of VMs deployed in the Cloud provider’s data center. Requests submitted by the users are provisioned and served if the Cloud has enough available capacity in terms of physical machines (PMs). Large Cloud service providers such as IBM provide service level agreements (SLAs) regulating the availability of the Cloud service. Before committing an SLA to the customers of a Cloud, the service provider needs to carry out availability analysis of the infrastructure on which the Cloud service is hosted. This chapter shows how stochastic analytic models can be utilized for Cloud service availability analysis. It first provides a background on the subject describing how the problem is faced in the current literature. Then, it proposes an example of one-level monolithic model that can be used to analyze the availability of an IaaS Cloud. However, such monolithic models become intractable as the size of Cloud increases. To overcome this difficulty, the chapter illustrates the use of an interacting sub-models approach. Overall model solution is obtained by iteration over individual sub-model solutions. Comparison of the results with monolithic model shows that errors introduced by model decomposition are negligible. It also shows how closed form solutions of the sub-models can be obtained and demonstrate that the approach can scale for large size Clouds. The presence of three pools of PMs and the migration of them from one pool to another caused by failure events makes the model both novel, interesting and particularly suitable in federated environments. In order to automate the construction and solution of underlying Markov models, the authors use a variant of stochastic Petri net (SPN) called stochastic reward net (SRN). This paradigm is supported by SHARPE (Trivedi & Sahner, 2009) and Stochastic Petri Net Package (SPNP) (Hirel, Tuffin, & Trivedi, 2000) software packages.
Rest of the chapter is organized as follows. Section II gives a background on the subject illustrating the state of the art. Section III provides an introduction to the formalism that will be used in the following to model the considered scenario. Section IV describes Cloud system model, assumptions and problem formulation. Section V, presents the monolithic SRN model. Interacting SRN sub-models are described in Section VI and their closed form solutions are presented in Section VII. Fixed point iteration among the interacting sub-models and proof of existence of a solution is shown in Section VIII. Results obtained from monolithic approach and interacting sub-models approach are compared in Section IX. Sections X and XI discuss how the approach can be used by Cloud providers, point out future challenges and highlight the benefit of decomposed models in the analysis of federation scenarios. The chapter concludes in Section XII.