Multi-processor parallel computer architecture using a parallel machine with topology-based mappings of composite grid applications

摘要

A parallelization process for complex-topology applications is based on an understanding of topology and includes two separate parts: i) an automatic, topology-based data distribution method and ii) a program transformation method. Together these methods eliminate the need for user determined data distribution specification in data layout languages such as High Performance Fortran. The topology-based data distribution method uses both problem and machine topology to determine a data-to-processor mapping for composite grid applications. The program transformation method incorporates statements in the user program to read and implement the data layout determined by the distribution method and to eliminate the need for user development and support of subroutine clones for data distribution.