An efficient algorithm to reduce the inflations in multi-supertask environment by using a transient behavior prediction method

The supertask approach is a means of supporting non-migratory tasks in Pfair (proportionate-fair) scheduling systems. In this approach, tasks bound to the same processor are combined into a single server task, the supertask, which is scheduled as an ordinary Pfair task. When a supertask is scheduled, one of its component tasks is selected for execution. P. Holman et al. showed that component-task deadlines can be guaranteed by inflating each supertask´s utilization. Their experimental results showed that the required inflation factors should be small in practice. Consequently, the average inflation produced by their rules is much greater than that actually required by the supertasks. First, we propose a notion of transient behavior prediction for supertasks, which predicts the latest possible finish time of subtasks that belong to supertasks. On this basis, we present an efficient schedulability algorithm for Pfair supertasks in which the deadlines of all component tasks can be guaranteed. We also propose a task merging process which combines the unschedulable supertasks with some Pfair tasks; hence, a new supertask can be scheduled in the system. Finally, we propose new reweighting functions that can be used when the previous two methods fail. Our functions produce a smaller inflation factor than previous work does. To demonstrate the efficacy of the supertasking approach, we present experimental evaluations of our algorithm, which decreases substantially the number of reweights and the size of inflation when there are many supertasks in the Pfair-scheduled systems.