HR: 0800h
AN: IN21B-0486    [Abstracts]
TI: The Rigorous Calculation of the Covariance Matrix for Arbitrarily Large Inverse Problems
AU: * Gunter, B C
EM: b.c.gunter@tudelft.nl
AF: Delft University of Technology DEOS - PSG, Kluyverweg 1, Delft, 2629HS, Netherlands
AU: van de Geijn, R
EM: rvdg@cs.utexas.edu
AF: The University of Texas at Austin Department of Computer Sciences, 1 University Station, Austin, TX 78712, United States
AB: The outline of an approach for computing the covariance matrix of a large, dense linear system will be presented. The creation of such matrices often arises in the Earth sciences, in particular when the error characteristics of a complex inverse problem are desired. For extremely large problems, such as those exceeding the memory limits of the machine being used, approximation techniques are often employed; however, doing so inherently introduces errors into the final covariance matrix. To avoid these errors, an approach has been developed by which the covariance matrix of a nearly arbitrarily sized dense linear problem can be computed on a machine of nearly any size. The concept involves the use of out-of-core computations, in which the problem is stored on disk and processed incrementally. While the idea of out-of-core computing has been around for decades, the specific approach we have developed involves the use of tiles as opposed to the more traditional slab-based approach. The tile-based method will be shown to provide both high performance and scalability on a range of problem sizes, number of processors, and machine types. A specific application involving the computation of the Earth's gravitational potential field will be illustrated.
DE: 0560 Numerical solutions (4255)
DE: 1200 GEODESY AND GRAVITY
SC: Earth and Space Science Informatics [IN]
MN: 2007 Fall Meeting