Algorithms to Harness Massive Parallelism.

摘要

This research project has been a study of many ways to control and to use massive parallelism in distributed computing systems. This research emphasized the use of hierarchies to simplify control and application algorithms intended for parallel computer systems of thousands or millions of processors. Most of the algorithms considered were organized to use interprocessor message paths that formed virtual spanning trees touching every node in a network. Simulators were developed to predict the collective behavior of hundreds or thousands of computers running local copies of the same algorithm; a graphical interface (PARVU) was developed to display moving color images of parallel program behavior in large networks; and fiber optic (MERLIN) hardware interfaces were designed to provide limited sharing of memory in networks of many hundreds computers running scientific application programs.