PERFORMANCE OF MASSIVELY-PARALLEL COMPUTERS FOR SPECTRAL ATMOSPHERIC MODELS

Massively parallel processing (MPP) computer systems use high-speed in terconnection networks to link hundreds or thousands of RISC microproc essors. With each microprocessor having a peak performance of 100 or m ore megaflops per second, there is at least the possibility of achievi ng very high performance. However, the question of exactly how to achi eve this performance remains unanswered. MPP systems and vector multip rocessors require very different coding styles. Different MPP systems have widely varying architectures and performance characteristics. For most problems, a range of different parallel implementations is possi ble, again with varying performance characteristics. In this paper, we provide a detailed evaluation of MPP performance for a spectral trans form kernel as used in weather and climate modeling applications. Usin g a specially designed spectral transform code, the authors study perf ormance on three different MPP systems: Intel Paragon, IBM SP2, and Cr ay T3D. Great care is taken to tune the implementation for efficient e xecution on each platform. The results yield insights into MPP perform ance characteristics, parallel spectral transform algorithms, and codi ng style for MPP systems. The authors conclude that it is possible to construct parallel models that achieve multigigaflops-per-second perfo rmance on a range of MPPs, if the models are constructed to allow comp ile- or run-time selection of some parallel implementation options.