Workload characterization of real time computing systems.
Michael H. Woodbury, Kang Geun Shin · Deep Blue (University of Michigan) · 1988
Proper workload analysis is often overlooked in performance and reliability studies of computing systems. Workload characterization is essential for the evaluation of any computer system, because system behavior is directly related to the type of workload it is h and ling. Based on the level of abstraction, the workload of a computing system is the collection of processing requirements presented to the system during a specified period of time. Despite the growing use and importance of real-time systems, their exclusive analysis has not been largely addressed in the literature relating to system performance and reliability. Significant results can only be obtained if the analysis is narrowed to the structure of a real-time system and its workload. This dissertation presents the analytic development and experimental justification of correlated mathematical models to study workload effects on performance and reliability for critical real-time systems. An underlying premise is that all theoretical results are experimentally verifiable. Workload analysis addresses different features depending on the level of abstraction. After defining a hierarchical modeling framework, the approach is to first analyze the workload at the sub-task level. We develop a probabilistic model involving the recursive use of task structure diagrams to model the structure of any real-time task. Based on reasonable assumptions, analytic expression for the active task time distribution are derived. The active task time is the total time a task is executing or waiting to execute. Once individual tasks are characterized, some multiple task effects are addressed. Thus, the second phase of research is modeling the structure and behavior of a complete real-time workload assuming that individual tasks have been determined and modeled. Given a scheduling policy, we show how to determine closed form expressions for task scheduling delay and active task time distributions. These results are used to determine the probability of dynamic failure and processor utilization. Finally, given that a complete workload is specified, two workload model applications are addressed. One is the workload effects on fault latency. This study discusses workload dependent performance effects on system reliability. The other study is the specification and development of a synthetic workload generator for real-time applications.