Automation of the application of data distribution with overlapping in distributed memory

The article deals with block-affine data layouts with overlapping for optimizing parallel computing in a distributed memory computing system. Examples of target computing systems are high-performance clusters and advanced systems on a chip with a large number of computing cores. It is proposed to describe the placement of an array with overlaps as a new array of slightly greater length, in which additional elements have the values of some elements of the original array. The possibility of developing an automatic transformation (by the compiler) of the usual allocation of an array in distributed memory into a new array containing overlaps is being considered. The proposed method is illustrated by a well-known numerical algorithm for solving the heat conduction problem.