An $\mathsf{EDF}$ -based restricted-migration scheduling algorithm for multiprocessor soft real-time systems

• Authors: James H. Anderson;Vasile Bud;Umamaheswari C. Devi
• Venue: Real-Time Systems
• Year: 2008

There has been much recent interest in the use of the earliest-deadline-first ( $\mathsf{EDF}$ ) algorithm for scheduling soft real-time sporadic task systems on identical multiprocessors. In hard real-time systems, a significant disparity exists between $\mathsf{EDF}$ -based schemes and Pfair scheduling: on M processors, the worst-case schedulable utilization for all known $\mathsf{EDF}$ variants is approximately M/2, whereas it is M for optimal Pfair algorithms. This is unfortunate because $\mathsf{EDF}$ -based algorithms entail lower scheduling and task-migration overheads. However, such a disparity in schedulability can be alleviated by easing the requirement that all deadlines be met, which may be sufficient for soft real-time systems. In particular, in recent work, we have shown that if task migrations are not restricted, then $\mathsf{EDF}$ (i.e. , global $\mathsf{EDF}$ ) can ensure bounded tardiness for a sporadic task system with no restrictions on total utilization. Unrestricted task migrations in global $\mathsf{EDF}$ may be unappealing for some systems, but if migrations are forbidden entirely, then bounded tardiness cannot be guaranteed. In this paper, we address the issue of striking a balance between task migrations and system utilization by proposing an algorithm called $\mathsf{EDF-fm}$ , which is based upon $\mathsf{EDF}$ and treads a middle path, by restricting, but not eliminating, task migrations. Specifically, under $\mathsf{EDF-fm}$ , the ability to migrate is required for at most M驴1 tasks, and it is sufficient that every such task migrate between two processors and at job boundaries only. $\mathsf{EDF-fm}$ , like global $\mathsf{EDF}$ , can ensure bounded tardiness to a sporadic task system as long as the available processing capacity is not exceeded, but, unlike global $\mathsf{EDF}$ , may require that per-task utilizations be capped. The required cap is quite liberal, hence, $\mathsf{EDF-fm}$ should enable a wide range of soft real-time applications to be scheduled with no constraints on total utilization.