Granularity management for a large-grain data flow multiprocessor
Amr Mohamed Zaky, Greg Negelspach · PPSC · 1995
Large-Grain DAta Flow (LGDF) multiprocessors provide a natural environment for executing computationally-demanding real time applications (e.g. signal processing). First, expressing the application as a task graph of previously defined primitives, eases the development process and exposes the natural parallelism in the application. Second, the low-overhead dynamic scheduling strategies used in LGDF multiprocessor provide a certain degree of fault tolerance-a crucial requirement in critical real time applications. We are studying the problem of improving the periodic execution schedules for such graphs by manipulating the computation and communication granularity. Our testbed is a model for a uniform memory access static data flow multiprocessor, e.g. the AN/UYS-2.