Modeling and performance control of Internet servers

The paper describes modeling and performance control of an Internet server using classical feedback control theory. We show that classical feedback control can leverage on well-known real-time scheduling results to resolve one of the fundamental problems in Internet-servers today; namely, achieving overload protection and performance guarantees in the presence of load unpredictability. The research is motivated by the increasing proliferation of a new category of Web-based services, such as online trading, banking, and business transactions, where performance guarantees are required in the face of unpredictable server load. Failure to meet desired performance levels may result in loss of customers, financial damage or liability violations. State-of-the-art Web servers are not designed to offer such performance guarantees. We show that control theory offers a robust solution to the server performance control problem. We demonstrate that a general web server may be modeled as a linear time-varying system, describe the equivalents of sensors and actuators in that system, formulate a simple feedback loop, describe how it can leverage on realtime scheduling theory to achieve timing guarantees, and evaluate the efficacy of the scheme on an experimental testbed using a real web server (Apache), which is the most popular Internet server today. Experimental results indicate that control-theoretical techniques offer a promising way of achieving desired performance in emerging critical Internet applications.