A Case Study in High-Performance Mixed-Language Programming

Abstract

Several widely used and promising programming tools and styles for computational science software are reviewed and compared. In particular, we discuss function/subroutine libraries, object-based programming, object-oriented programming, generic (template) programming, and iterators in the context of a specific example involving sparse matrix-vector products. A key issue in the discussion is to hide the storage structure of the sparse matrix in application code. The role of different languages, such as Fortran, C, C++, and Python, is an integral part of the discussion. Finally, we present performance measures of the various designs and implementations. These results show that high-level Python programming, with loops migrated to compiled languages, maintains the performance of traditional implementations, while offering the programmer a more convenient and efficient tool for experimenting with designs and user-friendly interfaces.