| Popis: |
Dynamic scheduling in real-time systems involves dynamically making a sequence of decisions concerning the assignment of system resources to real-time tasks. Tasks may have arbitrary time constraints, different importance levels, and fault tolerance requirements. Unfortunately, making these scheduling decisions is difficult, partly because the decisions must be made without the full knowledge of the future arrivals of tasks and partly because scheduling has to deal with many complex issues, e.g., multiprocessors and fault tolerance. Many existing algorithms, such as the Earliest Deadline First algorithm, cannot provide performance predictability for this dynamic scheduling problem. This dissertation presents a set of solutions, in the form of both theory and algorithms, with the goal of providing performance predictions and guarantees for three different scheduling problems. First, we provide a worst case analysis for algorithms that dynamically schedule independent tasks, where tasks are assumed to have different values. We derive performance bounds for both uniprocessor and dual-processor on-line scheduling. A set of threshold-based scheduling algorithms is found. These are shown to be better than the popular Earliest Deadline First algorithm. Secondly, for scheduling tasks with additional resources, we study dynamic scheduling algorithms based on the ability to generate feasible schedules and the quality of the generated feasible schedules, expressed in terms of the schedule length. Based on the analysis of two known algorithms, namely, list scheduling and the Spring heuristic scheduling algorithm, a set of new algorithms is developed and shown to possess better performance. Thirdly, for scheduling tasks with fault-tolerance requirements, assuming that reliability of task execution is achieved through task replication, we present an approach to mathematically determine the replication factor for tasks. The goal is to maximize the total performance index, which is a performance-related reliability measurement. We present a technique based on a continuous task model and show how it very closely approximates discrete models and tasks with varying characteristics. |
