A Formalism for Describing Data Distribution

In many existing and planned parallel machines, memory cannot be considered as a single homogeneous resource. Instead, each processor has a "local" section of memory which is more accessible than others. Because of this ease of access, it is necessary to distribute the data. across the system so that most references are made to local data. In this paper, we give a. mathematical description of data distribution in parallel machines. We then show its application to strip mining, a common transformation for converting sequential programs to run on parallel hardware. Strip mining using data distribution information enhances the locality of reference in the resulting program, thus speeding performance.