Task Merging for Better Scheduling

Abstract

AbstractThis paper proposes a new algorithm to restructure task graphs for suitable scheduling. The algorithm reduces communication costs by merging those tasks within a task graph whose communication costs exceeds their execution time. Task duplication techniques are applied before the merge, to avoid any delay in the execution of the tasks dependent on the merged tasks. Our experiments with a number of known benchmark task graphs demonstrate the distinguished scheduling results provided by applying our task merging algorithm before the scheduling.