The Case for Non-Preemptive Scheduling in Distributed Real-Time Systems

Contrary to traditional belief, we show in this paper, that for distributed systems non-preemptive scheduling can perform better than preemptive scheduling in the worst case in terms of task schedulability, under certain circumstances. We derive a worst-case delay bound for tasks scheduled using non-preemptive scheduling in a distributed system, where the task flow paths form a directed acyclic graph. The delay bound leads to a reduction of the distributed system to an equivalent hypothetical uniprocessor system scheduled using preemptive scheduling. This transformation enables the wealth of uniprocessor schedulability theory to be applied to analyze schedulability under non-preemptive scheduling in distributed directed acyclic systems. Our simulation studies show that non-preemptive scheduling can perform better than preemptive scheduling for distributed systems. We also characterize through simulations the scenarios under which non-preemptive is better than preemptive scheduling, and scenarios where the opposite is true. We hope this paper will serve as a first step towards more extensive study and use of non-preemptive scheduling in distributed systems.