Software engineering in parallel and distributed scientific computing: a case study from industrial practice

We report on results from an interdisciplinary research project of computer scientists, mechanical engineers and numerical analysts from industry and academia. We have designed and implemented a portable distributed memory parallel version of CFX-TfC-a state of the art computational fluid dynamics simulation software package that features finite volume discretization on unstructured hybrid grids in combination with an algebraic multigrid solver. Performance results on a number of parallel hardware platforms indicate high efficiency and scalability. Systematic application of software engineering methods has been a key contribution to success and helped to avoid bottlenecks found in many other parallel CFD codes. A complete re-design of CFX-TfC (which is implemented in Fortran 77) has been set up according to the object oriented paradigm. CFX-TfC´s key module, the algebraic multigrid solver (AMG), has been implemented in C++. The current run time efficiency is quite acceptable but could still be improved. Object oriented software engineering has considerably increased the solver´s maintainability. Since networks of workstations (NOWs) are of particular interest for small and medium enterprises (SMEs), we have implemented a resource manager that allows PVM applications such as Par-CFX-TjC to be executed in batch mode without interfering with interactive users. The resource manager features transparent checkpointing and process migration