A Methodology for Performance Predictions of Future ARM Systems Modelled in UML

Abstract

The increasing complexity and short product cycles drive developers of mobile systems to analyse the performance of systems before hardware prototypes are available. Therefore, it is necessary to predict application runtimes with the help of simulations of system models. Miscellaneous components and factors of mobile devices affect the performance, e.g. caches, buses etc. In order to predict the performance of new system designs already during early stages of development, models of the timing behaviour are necessary. We have developed a modular timing simulator for models of typical mobile systems which can be used to predict the runtime of applications on future systems. Since UML is the de-facto standard for software modelling and widely used, we use UML to specify the hardware of the system. In this way, the gap between hardware and software modelling may be closed and performance analysis of application and system design are tight closer. The UML system model consists of an architecture model and an instruction behaviour description. The architecture model describes the components of the system and the connections between them and the behavioural model specifies the timing of the processor instructions. These models are used to simulate different configurations of an ARM9 system. Traces from one configuration are used to predict the performance of another configuration. Predictions for an ARM 11 system with parallel pipeline units are made.