Simulation-trace-based component performance prediction

Abstract

Large software systems are composed of interdependent distributed components, some developed in-house, some commercially available, and others developed by the customer. The system architecture, that is, the components comprising the system and their interconnections, typically varies for each design. Performance prediction is important for such software, including helping designers to select better designs and helping them to adjust the software architecture for better performance. To predict the overall system performance, we must have performance data for each component. Performance modelling for reused components was reported elsewhere. This paper presents a technology and its accompanying tool suite to obtain performance models of new components in a formal software architectural design specification given in a communicating extended finite state machine (CEFSM) model. Performance data for such new components are not available and must be collected through simulation. Our technique includes three steps: component specification instrumentation; simulation; and component stochastic performance model derivation from simulation trace. We applied our technology to a telecom application to predict the performance of its new components. Combined with the performance models of reused components from previous work, we were able to predict performance of the entire architectural design.