CHARACTERIZING AND INTERPRETING PERIODIC BEHAVIOR IN COMPUTER SYSTEMS
Abstract
Periodic tasks are often present in information systems, such as UNIX daemons that check for work at fixed intervals and tasks that exchange "I'm alive" messages. Since the presence of such tasks can cause periodic behaviors in response times, service times, transaction rates and other performance measures, it is important to characterize periodic behavior, e.g., to develop accurate predictive models, or to assist with diagnosing performance problems. Herein, we characterize periodic behavior in a manner specific to information systems by using trapezoidal shapes to describe warm-up, steady-state, and cool-down phases in the execution of periodic tasks. Using measurements of workstation and mainframe computer systems, we compare this trapezoidal characterization with two conventional approaches: phase averages and sine-waves. Our results suggest that the trapezoidal characterization works well and only requires a modest number of parameters.