Using Supertasks to Improve Processor Utilization in Multiprocessor Real-Time Systems

We revisit the problem of supertasking in Pfair-scheduled multiprocessor systems. In this approach, a set of tasks, called component tasks, is assigned to a server task, called a supertask , which is then scheduled as an ordinary Pfair task. Whenever a supertask is scheduled, its processor time is allocated to its component tasks according to an internal scheduling algorithm. Hence, supertasking is an example of hierarchal scheduling. In this paper, we present a generalized “reweighting” algorithm. The goal of reweighting is to assign a fraction of a processor to a given supertask so that all timing requirements of its component tasks are met. The generalized reweighting algorithm we present breaks new ground in three important ways. First, component tasks are permitted to have non-integer execution costs. Consequently, supertasking can now be used to ameliorate schedulability loss due to the integer-cost assumption of Pfair scheduling. To the best of our knowledge, no other techniques have been proposed to address this problem. Second, blocking terms are included in the analysis. Since blocking terms are used to account for a wide range of behaviors commonly found in actual systems (e.g., non-preemptivity, resource sharing, deferrable execution, etc.), their inclusion is a necessary step towards utilizing supertasks in practice. Finally, the scope of supertasking has been extended by employing a more flexible global-scheduling model. This added flexibility allows supertasks to be utilized in systems that do not satisfy strict Pfairness. To demonstrate the efficacy of the supertasking approach, we present an experimental evaluation of our algorithm that suggests that reweighting may often result in almost no schedulability loss in practice. ∗Work supported by NSF grants CCR 9972211, CCR 9988327, and ITR 0082866.