作者： Henry Gabb , Bill Magro , Rudolf Eigenmann , Greg Gaertner , Steve Bova
DOI:
关键词: Parameter space 、 Parallel computing 、 Amplitude 、 Message passing 、 Sea state 、 System of linear equations 、 Finite element grid 、 Computer science 、 Thread (computing) 、 init
摘要: MPI and OpenMP were used concurrently in the CGWAVE application to improve performance of harbor analyses. The project, “Dual-level parallel analysis wave response using OpenMP,” won Most Effective Engineering Methodology award SuperComputing‐98 HPC Challenge competition. In this case was model motion waves Ponce Inlet on Florida’s Atlantic coast [1]. sea state a is characterized by number incident components, which are defined period, amplitude, direction. This set components can be regarded as parameter space; each triplet leads separate partial diffe rential equation solved finite element grid. Parallelism exploited at parameter-space level, with distribute work. Since execution times for differ much factor four, simple master‐slave strategy dynamically balances workload. For component calculation, large sparse linear system equations formed. solution parallelized OpenMP. OpenMP-based parallelization performed problem. Developing hybrid parallelism easier than expected, once few problems had been resolved. Problems occur mixed MPI/OpenMP program when MPI_Init or other communication routines called from an region. CGWAVE, we able avoid problem allowing only one thread execute message-passing operations.