The Real-Time Computing Laboratory, Software Predictability Group

Software Performance Control

Sponsor: NSF

Many modern applications, such as multimedia streaming, embedded real-time computing, reliable transaction processing, web hosting, and e-commerce require some form of performance guarantees. Two forms of performance guarantees have been investigated in prior literature; namely absolute and statistical. In this project, we investigate a different form we call convergence guarantees. These guarantees are softer than deterministic but offer more information on transient behavior than statistical.

In essence a convergence guarantee states that the system will converge to the desired specifications under certain conditions and within a specificed convergence envelope. Control theory is used as the underlying theoretic engine to provide convergence guarantees. The intuitive reason why performance of computing services can be modeled within a control-theoretic framework is because this performance depends on queueing delays at different computing resources (e.g., CPU, sockets, semaphores, network). Queues act as integrators of flows. Hence the computing system can be modeled by differential equations amenable to control-theoretic analysis.

Control systems have traditionally been employed in industry to manipulate the value of controlled system parameters. In particular, robust control and adaptive control theory deal with the problem of achieving a desired system performance in the absence of complete knowedge of controlled system dynamics. We are interested in applying elements of these theories in the Internet application domain to control the behavior of distributed performance-critical services in the absence of good system models and a priori information on load and resource usage patterns.

Software:

To provide one initial embodiment of the analytic framework, Ronghua Zhang, my former Ph.D. student (co-advised with Jack Stankovic), developed a middleware service called ControlWare for software performance control. It can be downloaded from the software link on his web-page.

Publications: