Abstract:
The accuracy of numerical weather prediction is essentially dependent on
the spatial resolution of the atmospheric model used for this purpose and on
the complexity and accuracy of the involved subgrid-scale
parameterizations (i.e., the parameterizations of processes whose typical
scale is smaller than the spatial step). A possibility of application of
models with higher spatial resolution and more precise parameterizations is
determined by the available computer technique and by the presence of
high-performing codes. Some trial versions of the atmospheric spectral model
with increased spatial resolution have been developed with the purpose to
improve the quality of numerical weather prediction. The corresponding code
was optimized and parallelized using the MPI technology. The code uses
one-dimensional decompositions over latitude in the physical space and over
spectral numbers in the spectral space. The developed parallel versions
of the spectral model (in different configurations) were tested and
the speed-up of the code was estimated as a function of the number of processors
for various computing platforms. Different methods of inter-processor data
exchange were examined and their efficiency was estimated.
This work was supported by the Russian Foundation for Basic Research
(porojects 04-05-64530a and 05-05-64575a).
The paper was
prepared on the basis of the authors' report at the International Conference
on Parallel Computing Technologies (PaVT-2007; http://agora.guru.ru/pavt).