Modeling and performance analysis of large scale IaaS Clouds

Abstract

For Cloud based services to support enterprise class production workloads, Mainframe like predictable performance is essential. However, the scale, complexity, and inherent resource sharing across workloads make the Cloud management for predictable performance difficult. As a first step towards designing Cloud based systems that achieve such performance and realize the service level objectives, we develop a scalable stochastic analytic model for performance quantification of Infrastructure-as-a-Service (IaaS) Cloud. Specifically, we model a class of IaaS Clouds that offer tiered services by configuring physical machines into three pools with different provisioning delay and power consumption characteristics. Performance behaviors in such IaaS Clouds are affected by a large set of parameters, e.g., workload, system characteristics and management policies. Thus, traditional analytic models for such systems tend to be intractable. To overcome this difficulty, we propose a multi-level interacting stochastic sub-models approach where the overall model solution is obtained iteratively over individual sub-model solutions. By comparing with a single-level monolithic model, we show that our approach is scalable, tractable, and yet retains high fidelity. Since the dependencies among the sub-models are resolved via fixed-point iteration, we prove the existence of a solution. Results from our analysis show the impact of workload and system characteristics on two performance measures: mean response delay and job rejection probability.

Introduction

An Infrastructure-as-a-Service (IaaS) Cloud, such as Amazon EC2 [1], IBM SmartCloud Enterprise, IBM SmartCloud Enterprise +[2], [3], and IBM Smart Business Desktop Cloud [4], delivers on-demand operating system (OS) instances provisioning computational resources in the form of virtual machines deployed in the Cloud provider’s data center. Such Cloud based services are gaining popularity leading to increasing business competitions and hence, performance and dependability guarantees are becoming critical. Providers of IaaS Clouds (e.g., IBM and Amazon) offer service level agreements (SLA) to Cloud users. Violations of such SLAs can cause loss of revenue and business reputation. We observe that, for most of the IaaS Cloud providers, offered SLAs are in terms of guaranteed availability. As technology and business models for Cloud services are getting mature, in future, users will also expect SLAs on Cloud performance besides availability. However, performance of an IaaS Cloud depends on a large number of factors including: (i) nature of physical infrastructure (e.g., CPU, RAM, disk characteristics), (ii) nature of virtual infrastructure (e.g., hypervisor characteristics), (iii) nature of management and automation tools (e.g., request provisioning steps), (iv) workload (e.g., arrival rate, request types) and (v) available capacity (e.g., number of physical machines). Hence, systematic performance assessment of Cloud infrastructure is difficult and non-trivial.

This paper presents a scalable analytic approach for model driven performance analysis of an IaaS Cloud. Traditional measurement based performance evaluation requires extensive experimentation with each workload and system configuration and may not be feasible in terms of cost due to the sheer scale of Cloud. A simulation model can be developed and solved but in contrast with an analytic model, it might be time consuming as the generation of statistically significant results may require many simulation runs [5]. A stochastic model is a more attractive alternative because of lower relative cost of solving the model while covering large parameter space. However, such stochastic analytic models are presumed not to scale well when dealing with the rising complexity associated with Cloud service architectures. Simplifying the model to make it more tractable can result in lower fidelity and, in the process, the effects of important parameters affecting the service performance metrics may not be captured [6]. To overcome this difficulty, in our recent research [7], we described a scalable approach using interacting stochastic sub-models where an overall solution is composed by iteration over individual sub-model solutions. We quantified effects of changes in workload (e.g., job arrival rate, job service rate) and system capacity (e.g., number of physical machines, number of virtual machines per physical machine) on service quality as measured by mean response delay and job rejection probability. In this paper, we extend our previous research in several directions:

(1) A monolithic Cloud performance model is constructed using our variant of stochastic Petri net (SPN) called stochastic reward net (SRN) [8]. We compare the scalability and accuracy of the interacting stochastic sub-models approach proposed in [7], w.r.t. the monolithic model. Our analysis shows that the monolithic model becomes intractable and fails to produce results as the scale of Cloud increases, while interacting sub-models approach quickly provides model solutions without significantly compromising the accuracy. Thus, we provide an analytic verification and validation of the proposed approach.

(2) Closed-form solutions of sub-models are shown whenever feasible. Such closed-form expressions can complement the use of analytic modeling software packages such as SHARPE [9] and SPNP [10], when dealing with large number of model states.

(3) Since dependencies among the sub-models are resolved via fixed-point iteration, in this paper, we prove the existence of a solution for the associated fixed-point equation.

(4) Numerical results are expanded and discussions are presented on how the proposed model can be extended to include different Cloud management aspects. Our developed model can be applied in capacity planning, forecasting, sensitivity analysis to find bottlenecks, what-if analysis or in an overall design optimization problem during design, development, testing and operational phases of an IaaS Cloud.

The rest of the paper is organized as follows. In Section 2, we present a system description, assumptions, and formulate the problems of interest. Section 3 describes interacting sub-models approach for performance analysis of IaaS Cloud. An equivalent monolithic model is described in Section 4. Numerical results are presented in Section 5. Generalizations of the interacting sub-models approach and future avenues of research are outlined in Section 6. Related research is highlighted in Section 7. Finally, we conclude this work in Section 8. Appendix A summarizes the symbols used in the paper and detailed steps of interacting sub-model closed form solutions are shown in Appendix B.