Topology Aware Performance Prediction of Collective Communication Algorithms on Multi-Dimensional Mesh/Torus

Abstract

As the scale of computers becomes large, technologies for appropriate selection of collective communication algorithm have become important. Especially, on Multi Dimensional Mesh/Torus topology, which is popularly used on large-scale supercomputers, this selection is one of the most important issues for achieving sufficient scalability. Currently, the authors are developing a method that combines performance prediction models and runtime measurement to enable efficient algorithm selection. This paper introduces a performance prediction model of algorithms on Multi-Dimensional Mesh/Torus topology. This model considers effects of collisions on links to increase the preciseness of the prediction. Some experiments have been done to show the advantages of the proposed model.