Kernel-control parallel versus data parallel

Abstract

The dusty deck problem is the problem of how to automatically transform large existing serial/vector codes into a form suitable for efficient execution on new parallel architectures. Kernel-control parallel (KCP) methods provide a promising approach to solving the dusty deck problem for MIMD distributed memory machines. This paper is a status report on a project to develop this technology into a full-scale prototype for running large Fortran 77 codes on the Intel iPSC/860. Space limits preclude a full description of the methods. Instead, we sketch the approach and then use a series of questions to explore some of the key differences between this new technology and existing data parallel methods, which are similar but better known.