The peer tasking design method

Abstract

This paper is a preliminary report of an ARPA sponsored study. It focuses on designing real-time command and control or battle management systems for parallel and distributed architectures. Due to delays in other ARPA programs, the targeted architectures were not available during the time frame of the study. The results of the study were, however, tested on more conventional sequential and parallel platforms. The design method discussed here is fundamentally different from those assumed by current real-time scheduling theories, e.g., rate-monotonic, earliest-deadline-first, least-laxity or best-effort. These theories assume that the fundamental unit of prioritization is the task. In this new method, the fundamental unit of prioritization is called a work item. Work items are functions the system performs that have timing requirements (deadlines) associated with them in the requirements specification. Current scheduling theories are applied using artifact deadlines introduced by designers whereas this new method schedules work items to meet specification deadlines (sometimes called end-to-end deadlines) required by the user. With this method, tasks have no priorities. A collection of tasks with no priorities will be called a collection of peer tasks. The study showed that it is possible to schedule work items based on importance rather than urgency while still meeting as many work item deadlines as can be met by scheduling tasks with respect to urgency. Also, it showed that the minimum on-line deadline that can be guaranteed for a work item of highest importance, scheduled at run-time, is approximately the inverse of the throughput, measured in work items per second, for a work load consisting only of work items of that type. Further, it was shown that it provides optimal utilization of a single processor machine, and that its timing behavior is predictable (provable) for both single and multiprocessor machines. Finally, it was shown that throughput is not degraded during overload.<>