A novel fault-tolerant scheduling algorithm for precedence constrained tasks in real-time heterogeneous systems

摘要

Fault-tolerance is an essential requirement for real-time systems, due to potentially catastrophic consequences of faults. In this paper, we investigate an efficient off-line scheduling algorithm generating schedules in which real-time tasks with precedence constraints can tolerate one processor's permanent failure in a heterogeneous system with fully connected network. The tasks are assumed to be non-preemptable, and each task has two copies scheduled on different processors and mutually excluded in time. In the literature in recent years, the quality of a schedule has been previously improved by allowing a backup copy to overlap with other backup copies on the same processor. However, this approach assumes that tasks are independent of one other. To meet the needs of real-time systems where tasks have precedence constraints, a new overlapping scheme is proposed. We show that, given two tasks, the necessary conditions for their backup copies to safely overlap in time with each other are (1) their corresponding primary copies are scheduled on two different processors, (2) they are independent tasks, and (3) the execution of their backup copies implies the failures of the processors on which their primary copies are scheduled. For tasks with precedence constraints, the new overlapping scheme allows the backup copy of a task to overlap with its successors' primary copies, thereby further reducing schedule length. Based on a proposed reliability model, tasks are judiciously allocated to processors so as to maximize the reliability of heterogeneous systems. Additionally, times for detecting and handling of a permanent fault are incorporated into the scheduling scheme. We have performed experiments using synthetic workloads as well as a real world application. Simulation results show that compared with existing scheduling algorithms in the literature, our scheduling algorithm improves reliability by up to 22.4% (with an average of 16.4%) and achieves an improvement in performability, a measure that combines reliability and schedulability, by up to 421.9% (with an average of 49.3%).