A hybrid parallel algorithm for the spectral transform method which uses functional parallelism

✍ Scribed by Edmund Chadwick

• Publisher: Elsevier Science
• Year: 1999
• Tongue: English
• Weight: 501 KB
• Volume: 25
• Category: Article
• ISSN: 0167-8191
• DOI: 10.1016/s0167-8191(98)00110-0

✦ Synopsis

A new hybrid parallelisation technique which combines functional parallelism with the more usual spatial parallelism is described and implemented for a global shallow water model of the earth's atmosphere using the Spectral Transform Method. Asymptotically, as the number of processors is increased, the functional parallelism shall be shown to have lower magnitude (than the spatial parallelism) for the number and volume of processor communications; also, the hybrid method shall be shown to give an extra freedom of choice in the parallelisation, as well as allowing a greater maximum number of processors for a given problem size; it is hoped that these advantageous factors will make the hybrid method a useful tool in the optimisation of parallel global models which use the Spectral Transform Method. The new technique is implemented for resolutions T85 and T170 on the Hitachi SR2201, a distributed memory RISC massively parallel supercomputer, and is found to reduce the execution times at certain resolutions.