Scheduling on the MasPar SIMD parallel computer

Abstract

This thesis studies the feasibility of a task scheduler for a parallel operating system. After analyzing several task scheduling algorithms, the highest level first algorithm was chosen. This algorithm has been empirically found to build schedules whose scheduling times are within 90% of optimal schedules 95% of the time. The highest level first algorithm was then modifies to handle the processor architecture on the Texas A&M University's MasPar parallel computer. These modifications included lowering communication costs by scheduling related tasks onto the same processor and assuming that processors may be dynamically allocated during task scheduling. Comparisons are made between serial and parallel versions of the task scheduler. The parallel version was faster at labeling tasks in the task graph, but the parallel version could not place the tasks onto processors in parallel. Nevertheless, the highest level first scheduling algorithm was acceptable for use in parallel operating systems.