Application Runtime Estimation for AURIX Embedded MCU Using Deep Learning

Estimating execution time is a crucial task during the development of safety-critical embedded systems. Processor simulation or emulation tools on various abstraction levels offer a trade-off between accuracy and runtime. Typically, this requires detailed knowledge of the processor architecture and high manual effort to construct adequate models. In this paper, we explore how deep learning may be used as an alternative approach for building processor performance models. First, we describe how to obtain training data from recorded execution traces. Next, we evaluate various neural network architectures and hyperparameter values. The accuracy of the best network variants is finally compared to two simple baseline models and a mechanistic model based on the QEMU emulator. As an outcome of this evaluation, a model based on the Wavenet architecture is identified, which outperforms all other approaches by achieving a mean absolute percentage error of only 1.63%.