Dependency-driven Parallel Programming

With the appearance of low-cost, highly parallel hardware architectures, software portability between such architectures is in great demand. Software design lacks programming models to keep up with the continually increasing parallelism of today’s hardware. This setting calls for alternative thinking in programming. When a computation has a static data-dependency pattern, extracting this pattern as a separate entity in a programming language, one can reformulate the computations. As a consequence, data-dependencies become active participants in the problem solving code. This allows us to deal with parallelism at a high-level. Data-dependency abstractions facilitate the mapping of computations to different hardware architecture without the need of rewriting the problem solving code. This in turn addresses portability and reusability issues.