Parallelizing a PDE solver: experiences with PISCES-MP

The paper presents a methodology for adapting dusty deck PDE solvers for parallel execution. Our approach minimizes changes to existing code and data structures, thereby preserving the value captured within dusty decks. This scheme uses the single program multiple data programming paradigm on message passing distributed memory architectures. To demonstrate the viability of our methodology the commercially available, dusty deck semiconductor device modeling program, PISCES, has been adapted for parallel execution. Simulating realistic complex device structures, we have achieved excellent performance gains over high performance serial workstations. Also, the scalability of the parallel simulator allows the simulation of structures too large for our existing serial computers.<
>