Assessing the Robustness of Arrival Curves Models for Real-time Systems

Design of real-time systems is prone to uncertainty due to software and hardware changes throughout their deployment. In this context, both industry and academia have shown interest in new trace mining approaches for diagnosis and prognosis of complex embedded systems. Trace mining techniques construct empirical models that mainly target achieving high accuracy in detecting anomalies. However, when applied to safety-critical systems, such models lack in providing theoretical bounds on the system resilience to variations from these anomalies. This paper presents the first work that derives robustness criteria on a trace mining approach that constructs arrival-curves models from dataset of traces collected from real-time systems. Through abstracting arrivalcurves models to the demand-bound functions of a sporadic task under an EDF scheduler, the analysis presented in the paper enables designers to quantify the permissible change to the parameters of a given task model by relating to the variation expressed within the empirical model. The result is a methodology to evaluate a system to dynamically changing workloads. We evaluate the proposed approach on an industrial cyberphysical system that generates traces of timestamped QNX events.