METHODOLOGY FOR EFFICIENT PERFORMANCE OF MULTISCALE MODELING METHODS IN HETEROGENEOUS HARDWARE INFRASTRUCTURES

摘要

The paper presents a novel approach dedicated to analysis and optimization of multiscale methods execution on heterogeneous hardware including processors, graphic cards and other accelerators. The methodology proposed in this paper is based on the investigation, which proved that efficiency of various numerical methods, used for computing purposes on different devices, often behaves unpredictably. This efficiency was analyzed in details and presented in this paper. On the basis of this assumptions two-step optimization procedure was proposed, i.e., before and during calculations, allowing to verify current efficiency and to propose new decomposition of computing domain. Both steps were described and verified in details by using loosely and fully coupled multiscale approaches. Performed validation with discussion of some unconsidered aspects is presented in the paper as well.