Transparent parallelisation through reuse: between a compiler and a library approach

Abstract

Software environments for commercially available Distributed Memory Parallel Computers (DMPCs) mainly consist of libraries of routines to handle communications between processes written in sequential languages such as C or Fortran. This approach makes it difficult to program massively parallel systems in both an easy and efficient way. Another approach relies on (semi-)automatic parallelizing compilers but it has its own drawbacks. We propose to tackle this problem at an intermediate level (i.e. between high level parallelizing compilers and raw libraries), using Object Oriented (OO) technologies. We show that existing OO techniques based on the reuse of carefully designed software components can be applied with satisfactory results to the large scale scientific computation field. We propose to use a form of parallelism, known as data parallelism, and to embed it in a pure sequential OOL (Eiffel). We illustrate on several examples how sequential components and frameworks can be modified for parallel execution on DMPCs to allow for transparent parallelisation of classes using these components and frameworks.