Dense linear algebra kernels on heterogeneous platforms: Redistribution issues

Abstract

Redistribution algorithms for dense linear algebra kernels on heterogeneous platforms are considered. In this context, processor speeds may well vary during the execution of a large kernel, which requires efficient strategies for redistributing the data along the computations. The proposed strategy is to redistribute data after some well-identified static phases and therefore is neither fully static nor fully dynamic. An optimal algorithm (under some assumptions) for redistributing data when computing the product of two matrices is presented.