Scheduling parallel program tasks onto arbitrary target machines

Abstract

We extend previous results for optimally scheduling parallel program tasks on a finite number of parallel processors. We introduce a new scheduling heuristic (MH) that schedules program modules represented as nodes in a precedence task graph with communication onto arbitrary machine topology taking contention into consideration. The results for scheduling simulated task graphs on ring, star, mesh, hypercube, and fully connected networks are also introduced. We also present Task Grapher, a tool for studying optimal parallel program task scheduling on arbitrarily interconnected parallel processors. Given a parallel program represented as a precedence-constrained task graph, and an interconnect topology of a target machine, Task Grapher uses one or more of its seven scheduling heuristics to produce the following displays: (1) Gantt Chart Schedule, (2) Speedup Line Graph, (3) Critical Path in Task Graph, (4) Processor Utilization Chart, (5) Processor Efficiency Chart, and (6) Dynamic Activity Display.