In this thesis we studied the problem of multiprocessor preemptive real-time scheduling taking into account the exact cost of the operating system (OS). This cost is composed of two parts: a part easy to determine, corresponding to the scheduler cost and another part difficult to determine, corresponding to the preemption cost. This difficulty is due to the fact that a preemption can involve another one, being able to so create an avalanche phenomenon. First, we studied the off-line multiprocessor real-time scheduling of independent tasks taking into account the exact preemption cost. We proposed a schedulability analysis based on a multiprocessor scheduling heuristic. This heuristic uses the partitioned multiprocessor scheduling approach. In order to take into account the exact preemption cost on every processor we use the schedulability condition proposed by Meumeu and Sorel. This schedulability condition for fixed priorities tasks, is based on a binary scheduling operation which counts the exact number of preemptions and add their cost in the schedulability analysis. The proposed heuristic maximizes the remaining utilization factor to fairly distribute the tasks on processors and to reduce their response time. It produces an off-line scheduling table. Secondly, we studied the off-line multiprocessor real-time scheduling of dependent tasks taking into account the exact preemption cost. Because the schedulability condition used for scheduling independent tasks can be applied only to fixed priorities tasks, it does not allow to manage priorities inversions that are involved by dependent tasks. We proposed a new schedulability condition for dependent tasks which enables fixed and dynamic priorities. This schedulability condition takes into account the exact preemption cost and dependences between tasks without any loss of data. Always with the partitioned scheduling approach, we proposed for dependent tasks a multiprocessor scheduling heuristic which reuses, on every processor, the schedulability condition proposed previously. In addition, this scheduling heuristic takes into account the interprocessors communication costs. It also minimizes on every processor the makespan (total execution time of the tasks on all the processors). This heuristic produces for every processor an off-line scheduling table. Supposing that we have a time-triggered multiprocessor architecture such that all the processors have a unique time reference, we proposed for every processor an on-line scheduler which uses the scheduling table produced during the off-line schedulability analysis. This on-line scheduler has the advantage to have a constant cost that is easy to determine exactly.Indeed, this cost corresponds only to the time necessary to read in the scheduling table the task selected for execution. In the on-line classical scheduler, this cost corresponds to the time necessary to update the list of ready tasks in order to select a task, according to a given scheduling algorithm, for example RM, DM, EDF, etc. In this case, the cost for selecting a task varies with the number of ready tasks which changes from an invocation of the scheduler to another one. Another advantage of the proposed on-line scheduler is that it is not necessary to synchronize the access to the data shared by several tasks, because this synchronization was already done during the off-line schedulability analysis.
