Towards cycle-accurate performance predictions for real-time embedded systems

In this paper we present a model-based performance analysis method for component-based real-time systems, featuring cycle-accurate predictions of latencies and enhanced system robustness. The method incorporates the following phases: (a) instruction-level profiling of SW components, (b) modeling the obtained performance metrics in MARTE- compatible models, (c) generation, schedulability analysis and simulation of a system model, (d) architecture improvement based on the analysis results. Our proposed method incorporates both the schedulability analysis and the simulation technique, complementing the advantages and eliminating the limitations of the individual steps. Moreover, the cycle-accurate performance metrics initiated by our method lead to accurate performance predictions for an autonomous navigation robot system, with only 6% deviation (or less) from the actual performance metrics