Date:

Monday , September 17, 2001

Speaker:

Tim Hoar

Affiliation:

National Center for Atmospheric Research, Geophysical Statistics Project

e-mail:

thoar@ucar.edu

Title:

Statistical Super Computing: formulas & flops

Abstract:

Many statistical applications in the atmospheric sciences are limited to small, "proof-of-concept" studies due to computational constraints. Traditional interpreted languages are either not available for todays fastest computers or simply do not take advantage of the architecture. We have implemented a spatio-temporal hierarchical Bayesian model to combine ocean surface wind data and a surface analysis field for the entire tropical Pacific. The data volume is quite large -- more than 700,000 bivariate observations every 4 days and it is necessary to predict at 24,576 locations every 6 hours (for as long as possible). The algorithm was developed in Matlab on a desktop and has been ported to F90 with MPI directives to run beyond the desktop, specifically an SGI Origin2000 and an IBM SP RS/6000 with 128 user-accessible nodes of 4 CPUs each.