Abstract. The paper is devoted to a distributed memory parallel implementation of a finite difference Eulerian global atmospheric model utilizing semi-implicit time stepping algorithm. The applied two-dimensional checkerboard partitioning of data in horizontal plane necessitates boundary exchanges between the neighboring processors all over the model code and multiple transpositions in the Helmholtz equation solver. Nevertheless, quite reasonable performance has been attained on a set of cluster multiprocessors.

The paper is focused on different approaches to parallel implementation of climate models on cluster multiprocessors. We present results of applying either MPI or OpenMP to the model components: atmospheric and oceanic blocks. A library of communication routines was developed for the distributed memory approach. It carries out the basic communication operations, such as boundary exchanges and transpositions of decomposed data. The library has quite general interface that makes it useful for parallelization of a wide range of scientific applications calculated on structured grids. In particular, we have examined it on INM AGCM as well as on a set of atmospheric benchmarks which integrate the equations of hydrothermodynamics of atmosphere under the hydrostatic assumption. An OpenMP implementation of the INM Ocean model has been undertaken and tested on few shared memory systems.