Atmospheric pollution transport: the parallelization of a transport & chemistry code

We discuss a model for computing the chemical reactions occurring among air pollutants and predict their dispersion in the atmosphere. We consider a parallel implementation on a CRAY T3D, using up to 128 processors. A regular subdomain decomposition is considered to optimize inter-processor communications, while load balancing – the main technical difficulty to obtain a good efficiency – is achieved by re-assigning temporarily pieces of work to other processors, according to a global strategy. The performance of our implementation is described by an analytical relation whose validity is checked on realistic data sets. A speedup of 100 is obtained with 128 processors.