Performance Analysis of a Large-Grain Dataflow Scheduling Paradigm

This paper describes and analyzes a paradigm for scheduling computations on a network of multiprocessors using large-grain dataj_ow scheduling at run time. The computations to be scheduled must follow a static flow graph, while the schedule itself will be dynamic (i.e., determined at run time). Many applications characterized by static flow exist, and they include real-time control and digital signal processing. With the advent of computer-aided software engineering (CASE) tools for capturing software designs in dataflow-like structures, macrodataflow scheduling becomes increasingly attractive, if not necessary. For parallel implementations, using the macro-data]qow method allows the scheduling to be insulated from the application designer and enables the maximum utilization of available resources. Further, by allowing multitasking, processor utilizations can approach 100 percent while they maintain maximum speedup. Extensive simulation studies are performed on _-, 8-, and 16-processor architectures that reflect the effects of communication delays, scheduling delays, algorithm class, and multitasking on performance and speedup gains.