Principia Mathematica, Inc.

4/3/2008

3:30pm-4:30pm

Parameter estimation is a key component of constructing models of natural systems. For example, the calibration of a groundwater model typically requires the estimation of parameters such as hydraulic conductivity which cannot be adequately established by direct measurements. This is a compute intensive task that benefits greatly from parallel processing.

PEST is a popular parameter estimation program that uses the Gauss-Levenberg-Marquardt algorithm to find optimal parameters in a model-agnostic framework. The algorithm requires many model runs to be made for the calculation of a Jacobian matrix, the size of which depends on the number of parameters being estimated, and a so-called lambda search to find a minimum along a search vector in parameter space.

BeoPEST is a parallel implementation of PEST designed to be used in an ad hoc cluster environment. BeoPEST allows the clustering of nodes running different operating systems and hardware architectures. The implementation allows new nodes to dynamically join and be dropped from the cluster, which accommodates nodes and network failures. The implementation is ideally suited for an environment where compute nodes become available after hours, but may not be reliable or may be withdrawn at any time.

The heterogeneous environment combined with problem-specific granularity also poses scheduling challenges which the implementation must address to achieve high efficiencies.

The implementation uses a lightweight master-slave style protocol that is suitable for use across a wide area network. This helps resolve security issues and also permits multiple compute clusters to be tied together. This talk discusses the implementation of BeoPEST, as well as practical issues surrounding parameter estimation of complex models.

Department of Computer Science

University of Colorado Boulder

Boulder, CO 80309-0430 USA

webmaster@cs.colorado.edu

University of Colorado Boulder

Boulder, CO 80309-0430 USA

webmaster@cs.colorado.edu