Gradient techniques for performance prediction and control in multitier systems

Chen, Shuyi

Permalink

https://hdl.handle.net/2142/29505 Copy

Description

Title

Gradient techniques for performance prediction and control in multitier systems

Author(s)

Chen, Shuyi

Issue Date

2012-02-01T00:49:45Z

Director of Research (if dissertation) or Advisor (if thesis)

Sanders, William H.

Committee Member(s)

• Abdelzaher, Tarek F.
• Gupta, Indranil
• Schlichting, Richard

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• gradient model
• performance prediction
• multitier system

Abstract

In recent years, the emergence of cloud computing, and software paradigms such as service-oriented architectures (SOAs), has led to the proliferation of Web services. One of the most important factors of the Web services is the end-to-end performance. It is important for the service provider to be able to predict and control the end-to-end performance of the Web services. However, as more and more Web services adopt a multitier architecture, and advanced techniques are used to improve the scalability and responsiveness, the problem of performance prediction and control for multitier systems is becoming more and more challenging. This motivates the search for techniques for performance prediction and control in multitier systems that power these Web services. Traditional techniques like queuing networks and machine learning have been applied in performance prediction and control. While these techniques work well in static systems, most of them fail to capture the dynamic systems as they are frequently reconfigured at runtime. The goal of the work in this thesis is to develop a runtime technique that can predict accurately the performance of general multitier systems, yet requires very little knowledge. In this thesis, we propose the gradient model technique. It is a runtime technique that combines simple first-order model, high-level system knowledge, and lightweight online measurements. It can be used with general multitier systems with little effort. We present the implementations and evaluations of three specific gradient models, the link gradient ,workload gradient and CPU gradient. Our results show that the gradient models can accurately predict the end-to-end performance of the multitier systems under a wide range of configurations. In addition, we demonstrate the practical usage of the gradient model in different scenarios, including per-transaction response time optimization and energy conservation. Finally, we present the design of the gradient model toolkit. The gradient model toolkit enables users to easily implement new gradient model and use the model in their own applications. Our vision is that, with the general gradient model methodology, and the toolkit we built, performance prediction and control for general multitier systems can be made simpler for system administrators and service providers.

Graduation Semester

2011-12

Permalink

http://hdl.handle.net/2142/29505

Copyright and License Information

Copyright 2011 Shuyi Chen

Owning Collections