An efficient numerical model of air pollution transport using parallel algorithms
by
Lance M. Leslie
Centre for Environmental Modelling and prediction, UNSW
Coauthors: Milton Speer (Bureau of Meteorology, Sydney)

A large research effort is being carried out by the Centre for Environmental Modelling and Prediction (CEMAP), University of New South Wales, in the development of a computationally efficient, real-time, air pollution analysis and prediction model, initially for the Greater Sydney area. The model has a number of components and sub-components. Optimisation of the numerical algorithms contained within these components is required if the real-time objective is to be achieved. The components include, first, an assimilation period of 12 to 24 hours. The aim here is to make best use of observational data scattered widely in space and time, leading up to the initial state (analysis) time at t = 0. A large number of analyses have been produced and it has become clear that for our applications a 4-dimensional variational data assimilation scheme provides the most accurate representation of the initial state. Details of the scheme, known as 4D-VAR, will be described as well as the extensive preconditioning required to achieve high convergence rates in the 4D-VAR scheme. The second component is the forecast model which itself has a number of sub-models including the advection model, the diffusion model and the UCLA air chemistry (photochemical transformations) model. Each of these sub-models requires optimisation.

Thus far, a series of experiments has been carried out on a set of 10 test cases, nominated by the NSW EPA as bad pollution days. The models have been run at a horizontal resolution of 1km on a 401 by 401 grid in the horizontal and 31 levels in the vertical, concentrated mostly in the lowest 300 hPa of the atmosphere. The results obtained from the simulations have used realistic archived atmospheric conditions but a largely synthetic chemical emissions inventory.

The air pollution model is a large-scale computational activity and, as such, requires high performance computing resources. To this end a parallel version of the model code has been implemented on a cluster of eight high performance workstations, in a conventional master/slave configuration. A simple decomposition of the computational domain into eight sub-domains has been employed. The sub-domains are overlapping to accommodate the advection and diffusion terms in the model, but non-overlapping for the model physics and the photochemical reactions. For the communications aspects of the model the standard Message Passing Interface (MPI) is used. Finally, the results of attempts at dynamic load balancing will be reported on.

Results obtained thus far indicate a speedup of approximately 5.7 for the eight-workstation cluster. Future work on further optimisation will be discussed.

Date received: July 21, 1999

Copyright © 1999 by the author(s). The author(s) of this document and the organizers of the conference have granted their consent to include this abstract in Atlas Conferences Inc. Document # cadk-35.