Performance specifications and metrics for adaptive real-time systems

While early research on real-time computing was concerned with guaranteeing avoidance of undesirable effects, such as overload and deadline misses, adaptive real-time systems are designed to handle such effects dynamically. Various research efforts have addressed the characterization and improvement of the dynamic behavior of real-time systems. However, to the authors' knowledge, no unified framework exists for designing adaptive, real-time software systems based on specifications of desired dynamic behavior. We propose such a framework based on control theory. Using control theory, a designer can (i) specify the desired behavior in terms of a set of performance metrics that can be mapped to a dynamic response of the control system, (ii) establish an underlying control model of the real-time systems, and (iii) design a resource scheduler using feedback control design methods to guarantee runtime satisfaction of the specifications. This is in contrast to more ad-hoc techniques. We also show that simply using long-term average performance metrics is not sufficient in designing controllers. We then develop a new algorithm based on two PID controllers that meet both the transient and steady-state performance requirements.