Scalable model of parallel computations for applications with intensive input-output

Abstract

A model of organization of computations on massively-parallel computer systems for applications with predominant input-output is presented. The model is constructed on the basis of distinguishing input-output patterns, redistribution of data and processing in data processing tasks. A strategy for separating the roles of supercomputer nodes is proposed. Features of the model implementation for systems with fat tree topology and three-dimensional torus topology are discussed.