Parallel Matrix Distributions: Have we been doing it all wrong?
Carter Edwards, Po Geng, Abani Patra, Robert A. Geijn · 1995
The basic premise of this report is that traditional matrix distributions for distributing matrices on distributed memory parallel architectures are in practice too restrictive. The primary problem lies with the fact that such distributions start with the matrix, not with the underlying physical problem. Through a series of examples, we show how this hampers convenient interfaces between applications and libraries. In some instances, we show how it hampers performance in general. We propose a new data distribution, Physically Based Matrix Distributions, which appear to show promise for solving the encountered problems. Some traditionally used distributions are shown to be a special, but often unnatural, case of this more general class of distributions. 1 Introduction Ever since the conception of distributed memory parallel computing, the problem of distributing data to the individual processors of a parallel computer has been of concern. Perhaps the longest studied problem has been th...