A simulation-based model generator for software performance estimation

With the rise of software system complexity, developers rely more on a modular approach to system design to reduce development cost. However, as a result, integrating a real-time system becomes a challenge. To be able to properly integrate the system, software developers are required to provide software characteristics such as the execution times of their components to ensure the correct timing behaviour of the overall system. Generally, engineers rely on profilers available on their workstations to collect execution times of software. Yet, the final target architecture is usually vastly different from that of the workstation. Furthermore, the fact that the target platform is mostly inaccessible at design time calls for tools that can estimate the execution time of components on a wide range of architectures with reasonable cost. In this paper, we propose a methodology that relies on (1) fast simulation techniques and (2) analytical tools that build predictive models to estimate the execution times of components on a target architecture with minimum detail. We show that the approach is able to predict the execution times of a set of benchmarks when migrated from a reference architecture to a target platform with comparable accuracy to simulation, while being 2 orders of magnitude faster.