1 Johannes Gutenberg University Mainz, Mainz, Germany
2 Tel Aviv University, Tel Aviv, Israel
Cloud-aerosol interactions play a key role in many aspects of cloud physics and chemistry such as the evolution of drop spectra, efficiency of rain formation, chemical characteristics of cloud and rain water, optical properties of the cloud and cloud processing of aerosols. The traditional treatment of detailed cloud microphysics in models are one dimensional distributions of drops and aerosol particles. This method has the advantage that it requires relatively short CPU time and moderate memory allocation. The disadvantage is that one dimensional distribution functions implicitely assume, that drops of the same size have the same contamination. This does not agree with observations: in atmospheric clouds, drops of the same size have often different chemical compositions, depending on the composition and amount of aerosols or gases dissolved in the drops. To overcome the shortcome stated above, multi-component distribution functions may be used (distribution functions that depend on the mass of water and on the masses of each of the aerosol species present). In comparison to one component distribution functions, multi-component ones provide a more realistic description of the physical and chemical processes in clouds. The disadvantage is the enormeous increase in computation time and the increased complexity in the numerical techniques. At the moment our work addresses the following two questions:
- How can we implement an affordable (in terms of computing time and memory requirements) numerical simulation of cloud-aerosol interactions?
- What is the relevance of the additional information we obtain by using a multi-dimensional distribution function?
A parallel version of a code, describing the condensational and collisional growth of a particle population was developed and implemented on massive parallel computers. The code parallelization has been done at the EPCC under TRACS funding and the model is presently running on a SP2 in Israel and a SPP2000 at Mainz. This presentation will give a short description of the algorithm and the implementation. The code parallelization was obtained by applying domain decomposition, using MPI. The computation is only carried out in those cells, which are occupied or are expected to become occupied during a time step. This causes a very inhomogeneous computational load. To overcome this problem, a dynamic domain decomposition has been developed, which is performed each timestep. In this presentation, the code performance on the HP X class-48 at Mainz is discussed. Apart from the speed-up obtained by the code parallelization in comparison to the sequential code, the load balancing problem resulting from the time evolution of the particle population and the resulting very inhomogeneous computational load are discussed. Examples of the time evolution of the particle distribution are shown and their physical relevance is discussed. The results obtained thusfar, allowed the following conclusions: A considerable code speed-up was obtained by the parallelization method applied. Eventhough, there are still improvements in the load balancing necessary. One method, which will probabably lead to a better load balancing would be the direct use of the computational load in each cell as load distribution factor instead of an occupancy function. In spite of the relatively simple numerical methods applied, it can be appreciated that the suggested model may provide a more realistic approach to treat cloud-aerosol interactions. During the collisional growth of the drops, mixing of drops of different contamination took place. Large drops of the same size still contained different amounts of aerosol particles.
1 Corresponding author address: Sabine C. Wurzler, Institute for Atmospheric Physics, Becherweg 21, Johannes Gutenberg University Mainz, D-55099 Mainz,