Ordonnancement des systèmes certifiés avec différents niveaux de criticité

Les systèmes temps-réels modernes ont tendance à obtenir la criticité mixte, dans le sens où ils intègrent sur une même plateforme de calcul plusieurs applications avec différents niveaux de criticités. D'un côté, cette intégration permet de réduire le coût, le poids et la consommation d'énergie. Ces exigences sont importantes pour des systèmes modernes comme par exemple les drones (UAV). De l'autre, elle conduit à des complications majeures lors de leur conception. Ces systèmes doivent être certifiés en prenant en compte ces différents niveaux de criticités. L'ordonnancement temps réel des systèmes avec différents niveaux de criticités est connu comme étant l’un des plus grand défi dans le domaine. Les techniques traditionnelles nécessitent une isolation complète entre les niveaux de criticité ou bien une certification globale au plus haut niveau. Une telle solution conduit à un gaspillage des ressources, et à la perte de l’avantage de cette intégration. Ce problème a suscité une nouvelle vague de recherche dans la communauté du temps réel, et de nombreuses solutions ont été proposées. Parmi elles, l'une des méthodes la plus utilisée pour ordonnancer de tels systèmes est celle d'Audsley. Malheureusement, elle a un certain nombre de limitations, dont nous parlerons dans cette thèse. Ces limitations sont encore beaucoup plus accentuées dans le cas de l'ordonnancement multiprocesseur. Dans ce cas précis, l'ordonnancement basé sur la priorité perd des propriétés importantes. C’est la raison pour laquelle, les algorithmes d'ordonnancement avec différents niveaux de criticités pour des architectures multiprocesseurs ne sont que très peu étudiés et ceux qu’on trouve dans la littérature sont généralement construits sur des hypothèses restrictives. Cela est particulièrement problématique car les systèmes industriels temps réel cherchent à migrer vers plates-formes multi-cœurs. Dans ce travail nous proposons une approche différente pour résoudre ces problèmes.Modern real-time systems tend to be mixed-critical, in the sense that they integrate on the same computational platform applications at different levels of criticality. Integration gives the advantages of reduced cost, weight and power consumption, which can be crucial for modern applications like Unmanned Aerial Vehicles (UAVs). On the other hand, this leads to major complications in system design. Moreover, such systems are subject to certification, and different criticality levels needs to be certified at different level of assurance. Among other aspects, the real-time scheduling of certifiable mixed critical systems has been recognized to be a challenging problem. Traditional techniques require complete isolation between criticality levels or global certification to the highest level of assurance, which leads to resource waste, thus loosing the advantage of integration. This led to a novel wave of research in the real-time community, and many solutions were proposed. Among those, one of the most popular methods used to schedule such systems is Audsley approach. However this method has some limitations, which we discuss in this thesis. These limitations are more pronounced in the case of multiprocessor scheduling. In this case priority-based scheduling looses some important properties. For this reason scheduling algorithms for multiprocessor mixed-critical systems are not as numerous in literature as the single processor ones, and usually are built on restrictive assumptions. This is particularly problematic since industrial real-time systems strive to migrate from single-core to multi-core and many-core platforms. Therefore we motivate and study a different approach that can overcome these problems.A restriction of practical usability of many mixed-critical and multiprocessor scheduling algorithms is assumption that jobs are independent. In reality they often have precedence constraints. In the thesis we show the mixed-critical variant of the problem formulation and extend the system load metrics to the case of precedence-constraint task graphs. We also show that our proposed methodology and scheduling algorithm MCPI can be extended to the case of dependent jobs without major modification and showing similar performance with respect to the independent jobs case. Another topic we treated in this thesis is time-triggered scheduling. This class of schedulers is important because they considerably reduce the uncertainty of job execution intervals thus simplifying the safety-critical system certification. They also simplify any auxiliary timing-based analyses that may be required to validate important extra-functional properties in embedded systems, such as interference on shared buses and caches, peak power dissipation, electromagnetic interference etc..The trivial method of obtaining a time-triggered schedule is simulation of the worst-case scenario in event-triggered algorithm. However, when applied directly, this method is not efficient for mixed-critical systems, as instead of one worst-case scenario they have multiple corner-case scenarios. For this reason, it was proposed in the literature to treat all scenarios into just a few tables, one per criticality mode. We call this scheduling approach Single Time Table per Mode (STTM) and propose a contribution in this context. In fact we introduce a method that transforms practically any scheduling algorithm into an STTM one. It works optimally on single core and shows good experimental results for multi-cores.Finally we studied the problem of the practical realization of mixed critical systems. Our effort in this direction is a design flow that we propose for multicore mixed critical systems. In this design flow, as the model of computation we propose a network of deterministic multi-periodic synchronous processes. Our approach is demonstrated using a publicly available toolset, an industrial application use case and a multi-core platform