SEMPA. Software Engineering Methods for Parallel Applications in Scientific Computing Final report

Methods for the parallelization of complex software systems in scientific computing have been developed and applied to the industrial CFD"6 simulation program TfC. TfC has been parallelized according to the SPMD"7 approach based on message passing with PVM or MPI. The parallel program, ParTfC, has been evaluated on a set of hardware platforms with industrial test cases; excellent efficiency and scalability has been shown for up to 128 processors. Using the algebraic multigrid solver, AMG, which is the numeric core of TfC, as a case study, we examined to what extent new programming languages and paradigms are applicable in the domain of scientific computing. Automatic or interactive parallelism tools as well as the dataparallel programming language High Performance Fortran are applicable only to a very limited extent to TfC because of TfC's irregular data structures. In contrast, object-oriented re-design of AMG and its implementation in C++ and Java has been successful. Moreover, a resource management system for networks of workstations has been developed that permits execution of production runs of parallel programs in batch mode by dynamically assigning resources that currently are not claimed by interactive users, to parallel jobs. It thus enables users in small and medium-size enterprises to use their existing hardware infrastructure as a platform for parallel calculations. (orig.)SIGLEAvailable from TIB Hannover: F98B1718 / FIZ - Fachinformationszzentrum Karlsruhe / TIB - Technische InformationsbibliothekBundesministerium fuer Bildung, Wissenschaft, Forschung und Technologie, Bonn (Germany)DEGerman