Statistical Analysis for Shared Resources Effects with Multi-Core Real-Time Systems

Today's multi-core and many-core COTS platforms make available a large amount of computational resource for real-time applications. As they aim at increasing performance for real-time, their challenges are the guarantees for timing constraints. Real time modeling and analysis are thus facing shared resources, optimization mechanisms, and sophisticated functionalities which all combine into complex system dynamics that are extremely costly to characterize. This paper proposes a measurement-based approach and a statistical analysis applied to define average and worst-case models to task executions under different possible execution conditions. The framework is formalized and then used to investigate different families of shared resources interference effects occurring on multi-core platforms; such effects are quantified with statistical metrics applied to measurements of tasks execution times. The focus of the work is on effects due to shared memories within the NXP T4240 multi core platform and the PikeOS hypervisor. A set of experiments is conducted to validate the framework proposed.