Multiprocessor fixed priority scheduling with limited preemptions

Abstract

Challenges associated with allowing preemptions and migrations are compounded in multicore systems, particularly under global scheduling policies, because of the potentially high overheads. For example, multiple levels of cache greatly increase preemption and migration related overheads as well as the difficulty involved in accurately accounting for them, leading to substantially inflated worst-case execution times (WCETs). Preemption and migration related overheads can be significantly reduced, both in number and in size, by using fixed preemption points in the tasks' code; thus dividing each task into a series of non-preemptive regions (NPRs). This leads to an additional consideration in the scheduling policy. When a high priority task is released and all of the processors are executing non-preemptive regions of lower priority tasks, then there is a choice to be made in terms of how to manage the next preemption. With an eager approach the first lower priority task to reach a preemption point is preempted even if it is not the lowest priority running task. Alternatively, with a lazy approach, preemption is delayed until the lowest priority currently running task reaches its next preemption point. In this paper, we show that under global fixed priority scheduling with eager preemptions each task suffers from at most a single priority inversion each time it resumes execution. Building on this observation, we derive a new response time based schedulability test for tasks with fixed preemption points. Experimental evaluations show that global fixed priority scheduling with eager preemptions is significantly more effective than with lazy preemption using link based scheduling in terms of task set schedulability

Response-Time Analysis of DAG Tasks under Fixed Priority Scheduling with Limited Preemptions

“…Moreover, they showed that the eager and lazy approaches are incomparable. A full schedulability analysis in the case of eager preemptions and multiple non-preemptive regions has been recently proposed in [15].…”

“…In [15], the authors derived the lower-priority interference that a sequential task can suffer due to lower-priority tasks, considering global fixed-priority scheduling with eager preemptions:…”

“…After τ k started executing, it could be blocked again by at most m − 1 lower priority tasks at each preemption point. Therefore, for sequential task-sets, the lowerpriority interference is upper-bounded considering: (1) the set of the longest NPR of each lower-priority task and then (2) the sum of the m and m − 1 longest NPRs of this set, as computed in [15]. This no longer holds for DAG-based task-sets, because multiple NPRs from the same task can execute in parallel.…”

“…In [15], the authors upper-bounded the lower-priority interference in a multi-core system executing sequential tasks by considering the longest NPR of lowerpriority tasks. However, in parallel DAG-based task-sets, multiple NPRs from different tasks can execute in parallel on different cores.…”

“…Our analysis, which builds upon [10] and [15], determine the interferences and blocking impact of higher-and lower-priority tasks respectively. In order to derive the lower-priority interference, we propose two methods: (1) a pessimistic but easy-to-compute method that upper-bounds the blocking impact based on the task-set's longest NPRs (named LP-max); and (2) a tighter but computationallyintensive ILP-based method (named LP-ILP) that analyzes which NPRs can actually execute in parallel to refine the blocking estimation.…”