The Rainman data set is a collection of Bureau of Meteorology ground
station measurements of rainfall for the whole of Australia, covering
approximately the past sixty years and four thousand collection points.
This data set can be used as raw data to an interpolation technique to
provide rainfall surface estimations.
Surface interpolation of irregularly positioned data points to a uniform grid enables us to combine this data with earth observation data such as the GMS-5 geostationary meteorological satellite and the NOAA polar orbiting satellite.
This work is being done in collaboration with the Soils and Land Management Cooperative Research Centre.
Kriging Interpolation is an interpolation technique well suited to data sets that exhibit clustering like the Rainman data set and involves solving a matrix problem set up around the known rainfall values. This technique requires frequent intensive matrix multiplications and matrix inversion, hence we need to use specialised matrix solvers.
We have developed
a Kriging routine which runs on the 128 node Connection Machine CM5 at the
South Australian
Centre for Parallel Computing
and is implemented using CMFortran and the CMSSL library which contains
optimised matrix solvers.
To contrast and explore the performance requirements of
the technique we have also implemented the Kriging technique using High
Performance Fortran on our farm of DEC AlphaStations.
This problem is a good example of an application where
selection of the execution platform dependent on application
parameters and data size is very important. We can explore the
tradeoff between the matrix multiplication and matrix inversion
costs depending on data input sizes and by modeling this tradeoff and
gathering metrics we can make a sensible selection per execution
request.
For more information, contact
Katrina Kerry.