A Method for Improving Efficiency of Static Program Graph Scheduling

Obtaining maximum performance from parallel distributed memory machines depends on partitioning parallel programs into modules and scheduling those modules for the shortest possible execution time. The scheduling problem is known to be NP-hard and heuristic algorithms have been proposed to obtain optimal and sub optimal solutions. The partitioning algorithm partitions an application into tasks with appropriate grain size and represents them in the form of a directed acyclic graph. The nodes of the resulting graph are then scheduled onto the processors of a parallel computer system. We can see that almost all coarse grained program graph nodes don’t need all of their input operands at the beginning of their execution. Thereafter, they can be scheduled a bit earlier. This type of program graph nodes triggering is called partial strict triggering. The missing operands will be requested later, during the execution. Coarse grained program graph nodes send their output operand to all successors, as soon as they produce them. Successors of coarse grained program graph nodes will be scheduled earlier too, because they will receive their input operands sooner. An evaluation of improved CPM, VL and DSH scheduling algorithms is done in this paper. We have improved them with partial strict triggering of coarse grained program graph nodes.