|
Atmospheric numerical models are used for applications of great socioeconomic importance including short-term weather prediction, future global climate projections, and seasonal predictions. The sophistication and spatial resolution of atmospheric models are subject to available computing power. Increased computing capabilities will allow global resolution of regional features for climate models, improving the large-scale flow via smaller-scale detail. For weather prediction models, crucial phenomena like convection could be simulated directly without uncertain parameterizations which estimate the net effects of unresolved phenomena.
Several high-end computers have finally breached the "Petaflop barrier" performing one quadrillion operations every second. This unprecedented computing power, however, is not being realized by current models leaving the effectiveness unduly limited. The main reason they are not realizing available power is communication burden which is very costly in a massively parallel environment. Spectral element methods in particular are on their way to operational climate models, and they effectively utilize current machines. Time steps for these methods tend to be smaller than often desirable though.
I am developing scalable explicit finite volume methods for atmospheric simulation which reduce the communication burden while allowing a larger time step than most explicit methods. These methods rely heavily on characteristic theory and share similar approaches as Constrained Interpolation Profile (CIP) methods. Currently, a 2-D non-hydrostatic atmospheric model for buoyancy-driven flows runs stably and accurately up to CFL=2. I am currently improving accuracy at larger CFL numbers.
I run these codes on Graphical Processing Units (GPUs) which I have found to be extremely effective, obtaining >20x speed-up over dual-core OpenMP CPU code using a single GPU in double precision with little to no memory tuning. GPU cost to performance ratio is an order of magnitude better than CPUs if the code efficiently utilizes GPUs.
Semazzi, Fredrick H. M., J. S. Scroggs, G. A. Pouliot, A. L. Mckee-Burrows, MATTHEW R. NORMAN, V. Poojary, and Y. M. Tsai: 2005. On the Accuracy of Semi-Lagrangian Numerical Simulation of Internal Gravity Wave Motion in the Atmosphere. Journal of the Meteorological Society of Japan, 83, 851-869.
NORMAN, MATTHEW R. and R. D. Nair, 2008. Inherently Conservative Non-Polynomial Based Remapping Schemes: Application to Semi-Lagrangian Transport. Monthly Weather Review, 136, 5044-5061.
NORMAN, MATTHEW R., F. H. M. Semazzi, and R. D. Nair, 2009. Conservative Cascade Interpolation on the Sphere: An Intercomparison of Various Non-Oscillatory Reconstructions. Quarterly Journal of the Royal Meteorological Society, in review.
|
