Predicting worst-case execution time trends in long-lived real-time systems

Abstract

In some long-lived real-time systems, it is not uncommon to see that the execution times of some tasks may exhibit trends. For hard and firm real-time systems, it is important to ensure these trends will not jeopardize the system. In this paper, we first introduce the notion of dynamic worst-case execution time (dWCET), which forms a new perspective that could help a system to predict potential timing failures and optimize resource allocations. We then have a comprehensive review of trend prediction methods. In the evaluation, we make a comparative study of dWCET trend prediction. Four prediction methods, combined with three data selection processes, are applied in an evaluation framework. The result shows the importance of applying data preprocessing and suggests that non-parametric estimators perform better than parametric methods.