Estimation of an Early WCET Using Different Machine Learning Approaches

Abstract

The Worst-Case Execution Time (WCET) is crucial and an essential factor in analyzing and developing Real-Time Embedded Systems. The idea of the WCET allows the designer to create safe and reliable real-time systems. The WCET is used by the scheduler to determine an appropriate scheduling scheme for the application and to guarantee timing constraints. These systems need to satisfy a strict deadline, and failing leads to catastrophic events such as loss of life. Generally, WCET analysis is applied only in the late stages of Safety-Critical Systems development when the hardware is available, and code is compiled and linked. Different methods exist to determine WCET, but none of these provide early insight into WCET. Suppose early WCET is unavailable during system development. In that case, many systems design decisions are made using experience, which might be impractical if systems don’t satisfy timing constraints, and it may result in costly system re-design. However, we need early WCET in the initial stages of systems development as an essential prerequisite to configure the system properly. We propose a method to determine early WCET using machine learning models without the need for the hardware and binaries, i.e., source-level timing analysis-this new approach estimate WCET using source code. The models can predict WCET without running the code on the platform once it is trained, which is created using the Pytorch framework. The feasibility of this approach is evaluated using six different machine learning models such as Support Vector Regression (Linear kernel and RBF kernel), Tree Regression, Forest Regression, K-Nearest Neighbors Regression and Ridge Regression. The viability of the proposed method is demonstrated with the TACLeBench benchmark suite. The results show reasonable estimates on the predicted execution time of the final and compiled code, proving the prospectus of the methodology.