2. Introduction

The regional climate model used in this study and called REMO in the following is a meso-a-scale grid-point model for numerical climate prediction derived from the EM/DM ("Europa-Modell", "Deutschland-Modell") of the German Weather Service, a model for numerical weather forecasting. For the parallelization we therefore have to solve almost the same problems as in parallelization approaches used in numerical weather prediction systems.

Above all these problems are:

• Organization of parallel input/output processing.

• Development of effective parallel implementations for the explicit Eulerian time-step, the semi-implicit corrections and the physical parametrizations based on a data parallel domain decomposition approach.

• Avoidance of load imbalances.

The original serial version of REMO is running on a shared memory vector architecture (Cray-C90) and is written in Fortran 77 code. This version of REMO is already parallel in the sense that it works under the Cray macrotasking control. The parallel version for distributed memory computers has been developed and optimized on an IBM RS/6000 SP (69 Thin nodes (66 Mhz POWER 2) with 128 MB RAM, 3 Wide nodes (66 MHz POWER 2) with 1 GB RAM, High-Performance Switch (40 MB per seconds, 39 usec latency), TB2 Adapter). The implementation was carried out by using Fortran 90 coupled with the Message Passing Interface (MPI) to create a portable parallel version which is also suitable for execution on other parallel platforms. Fortran 90 was chosen to ensure an efficient dynamic memory management.