Dense Linear Algebra Kernels on Heterogeneous Platforms: Redistribution Issues

In this paper, we deal with redistribution issues for dense linear algebra kernels on heterogeneous platforms. In this context, processors speeds may well vary during the execution of a large kernel, which requires efficient strategies for redistributing the data along the computations. The strategy that we propose is to redistribute data after some well identified static phases and therefore, it is neither fully static nor fully dynamic. We present an optimal algorithm (under some assumptions) for redistributing data when performing matrix matrix multiplication