A New Hybrid Approach on WCET Analysis for Real-Time Systems Using Machine Learning

The notion of the Worst-Case Execution Time (WCET) allows system engineers to create safe real-time systems. This value is used to schedule all software tasks before their deadlines. Failing these deadlines will cause catastrophic events, e.g. vehicle crashes, failing to detect dangerous anomalies, etc. Different analysis methodologies exist to determine the WCET. However, these methods do not provide early insight in the WCET during development. Therefore, pessimistic assumptions are made by system designers resulting in more expensive, overqualified hardware. In this paper, an extension on the hybrid methodology is proposed which implements a predictor model using Machine Learning (ML). This new approach estimates the WCET on smaller entities of the code, so-called hybrid blocks, based on software and hardware features. As a result, the ML-based hybrid analysis provides insight of the WCET early-on in the development process and refines its estimate when more detailed features are available. In order to facilitate the extraction of code-related features, a new tool for the COBRA framework is proposed. This paper proves the potential of the ML-based hybrid approach by conducting multiple experiments based on the TACLeBench on a first prototype. A set of annotated code features were used to train and validate eight different regression models. The results already show promising estimates without tuning any hyperparameters, proving the potential of the methodology.