Exploring SW performance using preemptive RTOS models

With increasing SW content of modern SoC designs, modeling of embedded SW has become critical. For one, analyzing software performance early in the system design flow is now paramount to an efficient implementation. Previous work addressed performance modeling with timing annotated functional models and exposed dynamic scheduling effects with behavioral RTOS models. However, such models insufficiently capture preemption as their cooperative decision making is dependent on the timing annotation granularity. In addition to capturing dynamic scheduling, modeling system overhead (e.g. for context switches) becomes essential for guiding developers when deciding the granularity of multitasking applications. In this paper, we introduce two means to improve accuracy of SW performance modeling: a preemptive RTOS model, and the modeling of system overhead. Our experimental results on multimedia applications significant accuracy improvements when analyzing interrupt latency distribution (within 8% for average and 50th percentile), and modeling systems with high system overhead (less than 10% error). Our model extensions provide improved simulation accuracy and therefore better aid the design space exploration.