Performance results of a large-scale air pollution model on two parallel computers

Abstract

The size of the computational task obtained after the appropriate splitting and discretisation procedures of the system of partial differential equations in large air pollution models is enormous. Even when computers with several GFLOPS top performance are in use, it is difficult to solve this problem efficiently and, furthermore, to prepare codes that can be used for operating purposes. Some performance-comparison results obtained by runs of the parallelised via MPI unified variable-grid version of the Danish Eulerian model for long-range transport of air pollutants on two parallel computer architectures are presented. The test numerical experiments are performed on a Beowulf-like cluster set-up of four two-processor Macintosh G4 machines and a multiprocessor SUN (up to 32 processors). The Message Passing Interface (MPI) standard is used on both distributed and shared memory parallel computer platforms. Timing results, speed-up and parallel efficiency of the main computational stages as well as of the whole model are presented.