Performance Optimization of an Atmospheric Model in Massively Parallel Computers

摘要

This paper discusses optimization issues to be addressed in producing a major imporvement in the performance of a state-of-the-art general circulation model of the atmosphere on massively parallel computer systems. Our work is based on the UCLA Atmospheric General Circulation Model (AGCM). The UCLA AGCM is a "grid-point model", which uses a three dimensional straggered grid to compute flow variables around the globe. The current goal is to produce a code that runs at the 50 GFLOP level by early 1998.