Performance-Driven Design for Control and Scheduling in Real-Time Systems

The development of most control applications in real-time embedded systems is based on the assumption that the design, definition and analysis of the control functionality can be separated from the development and analysis of the software code implementing it. The functional model of the controller and the model of the program threads are connected by a suitable set of assumptions on the time properties of the code. In most cases, these assumptions concern the activation periods, the worst case response times (compared to the deadlines) and possibly the output jitter. By following the standard approach, if all the tasks are schedulable within the deadlines, then the system is assumed to be correct. This assumption corresponds to the classic hard deadline scheduling model. In reality, a wide range of physical control systems exist, in which deadlines can be missed without catastrophic consequences. Restricting the controller period to be greater than the worst case response time - thus avoiding possible overruns - is safe, but often this maps to unnecessarily conservative design constraints. Relaxing those timing constraints opens the doors to a wider design space that potentially allows to significantly increase the control performance. This however requires a proper modeling { more accurate than the available state-of-art ones { to analyze the effects of missed deadlines on the system behavior. Following this novel approach, the problem should be treated as a co-design optimization, where the objective is to find the controls implementations and the scheduling policy in such a way that the performance of the whole system is optimized. This approach is extended also to the problem of porting those applications to modern multicore platforms in an optimal way, while guaranteeing all functional dependencies. Motivated by these challenges, this thesis contributes in building a performance-driven design approach that merges both the aspects of the physical system to be controlled and those coming from the platform architecture. We explore new models and new analysis techniques that study the impact of schedulability strategies on the performance of control systems, taking also into account the possibility of (sporadic) deadline misses. The Logical Execution Time paradigm, which has gained attention in recent years as a viable candidate for enforcing determinism in multicore applications, is here adopted and improved, to preserve causality and time determinism in all operating conditions. In particular, we provide methods and tools to study the effects on weakly-hard bounds when using different strategies to handle deadline misses; improve the design of control tasks that are robust to deadline misses; provide new ways to monitor the evolution of the control performance under non-ideal timing conditions; develop a framework to study these effects on complex cyber-physical systems via co-simulation; integrate new functionalities and legacy code on multicore platforms in an efficient way, while maintaining the correct functional behavior.