Suitability of GCM Physics for Execution on SIMD Parallel Computers

Abstract

A possible path to the teraflop performance that will be required by the next generation of GCMs (general circulation models) is the SIMD (Single Instruction stream, Multiple Data stream) massively parallel computer architecture. A critical consideration in moving a model to a SIMD architecture is the efficiency of the model's physical parameterizations on this type of machine. We have analysed the physics in a production GCM to evaluate its potential for efficient SIMD execution, on the assumption that each processor of the computer is allocated to a single vertical column of the model. This paper summarizes the results obtained for the various physics routines and compares these with the efficiencies obtained under the assumption of MIMD (Multiple Instruction stream, Multiple Data stream) execution. Overall, we found a performance penalty of only 15% to 20% for SIMD compared to MIMD execution. This is a very acceptable result, which suggests that SIMD computers should not be rejected for climate modelling (or numerical weather prediction) applications.