Institut für Organisation und Management von Informationssystemen

Zusammenfassung: Parallel and portable development of programs is an important subject for future computing systems. With the introduction of heterogeneous many-core systems and the increasing impact of the memory wall, classical software develop¬ment paradigms no longer hold true. Data access and communication cost pose a crucial performance obstacle. Dataflow models are therefore getting attention from large scale and performance oriented developers. Such models are principally error prone, and require reorganization of the code to maximize usability. This paper outlines a programming approach that annotates the code with algebraic logic, thus providing data dependencies information. This not only allows more intuitive and fine-granular programmability, but in particular enables code rearrangement for manipulating the execution behavior, its concurrency and even its destination platform. This model is specifically geared towards mathematical problems and thus addresses in particular High Performance Computing needs.