Libraries for parallel paradigm integration

A programming paradigm is a method for structuring programs in order to reduce the complexity of the programming task. In parallel programming, task and data parallelism are the most common paradigms, although object, functional, logic, and database parallelism are also used. Most existing programming languages and tools are designed to support either task parallelism or data parallelism. Yet there are many applications that can benet from a combination of both paradigms. For example, image processing problems can exploit the regularity of data-parallel computation within each stage of a pipeline. This process requires tools that support both task and data parallelism and that allow task- and data-parallel components to interact. In particular, it should be possible to reuse libraries developed in either paradigm in a variety of situations. To meet these requirements, we have designed a small set of extensions to Fortran, called Fortran M, that provides a framework on which portable, reusable libraries can be built to implement programming paradigms. In this paper, we illustrate the approach by showing how it is used to implement libraries for le I/O, message-passing, and data-parallel computation. 1