Field-Object Design of a Numerical Weather Prediction Model for Uni- and Multiprocessors

Abstract

The performance of a numerical weather model constructed from scalar, vector, and tensor field objects is evaluated on several workstation computers and on a message-passing multiprocessor. The motivation is to achieve high performance in mesoscale modeling at a low cost in both hardware and programming effort, by making models easier to develop and maintain and by simplifying parallel implementation. The design of this model helps achieve these goals by providing portability of Fortran or C++ code between workstations and fast message-passing multiprocessors and by coding with abstract datatypes to mimic a mathematical description of the model. The overhead of calling operators on field data types was found to be insignificant. Rather, performance on workstations was limited by high cache-memory turnover due to repeated applications of simple operators on an entire 3D grid, and this cache turnover rate was limited in turn by the memory subsystems of the workstations tested. However, parallel processing is a potential cure for this problem, since, for a given task, more processors mean less data through each cache and, hence, less turnover of cache data. Near linear speedup of the model task was observed on a small message-passing multiprocessor.