Introduction of deep neural network in hybrid wcet analysis

Abstract

Safe and responsive hard real-time systems require the Worst-Case Execution Time (WCET) to determine the schedulability of each software task. Not meeting planned deadlines could result in fatal consequences. During development, system designers have to make decisions without any insight in the WCET of the tasks. Early WCET estimates will help us to perform design space exploration of feasible hardware and thus lowering the overall development costs. This paper proposes to extend the hybrid WCET analysis with deep learning models to support early predictions. Two models are created in TensorFlow to be compatible with our COBRA framework. The framework provides datasets based on hybrid blocks to train each model. The feed-forward neural network has a high convergence rate and is able to learn a trend in the features. However, the error of the models are currently too large to predict meaningful upper bounds. To conclude, we summarise the problems we need to tackle to improve the accuracy and convergence issues.