next up previous
Next: About this document

Grand Challenge Applications and Enabling Scalable Computing Testbed in Support of High Performance Computing:
Title: Four Dimensional Data Assimilation
Milestone 10: 200-fold speedup for the Lagrangian filter over the equivalent problem solved using an Eulerian filter on a single node of a Cray C90

Agreement Number: NCCS5-150

January 26, 2000

Principal Investigator: P.M. Lystergif
Earth System Science Interdisciplinary Center
Department of Meteorology
University of Maryland College Park
and NASA/GSFC Data Assimilation Office

Abstract:

This is the documentation which is being provided in support of the following milestone submission for the project: Milestone 10: 200-fold speedup for the Lagrangian filter over the equivalent problem solved using an Eulerian filter on a single node of a Cray C90 (delivered with scaling analysis).

We have developed a new numerical algorithm, the Lagrangian filter, for solving the Kalman filter equations for constituent assimilation of observations using a direct solution on trajectories that propagate with the wind flow. This is a finite dimensional approximation of the solution by characteristics of the estimation problem and may be thought of as an extension of the well known method of trajectory mapping. We assert that the Lagrangian filter is a more natural framework for the study of the constituent problem with data assimilation than the equivalent Eulerian filter because of the conservative properties of field, error variance, and error covariance dynamics. Considerable insight in the behavior of the filters was gained as a result of these properties. The Lagrangian filter also requires fewer floating point operations than the Eulerian filter because of the simple forecast error propagation step. However it is still computationally expensive and we implemented it for two-dimensional flow in the stratosphere and validated it against the Eulerian filter. This code was developed using the distributed-memory parallel methodology -- in particular it uses the Message-Passing Interface (MPI) library. This work was enabled by the use of parallel computers under NASA's High Performance Computing and Communications (HPCC) program, and we have achieved on an SGI/Cray T3E-1200 computer a 200-fold speedup over the equivalent problem solved using the Eulerian filter on a single processor of a Cray C90.

The following is a link to a paper A Lagrangian Trajectory Filter for Constituent Data Assimilation by Lyster et al. ftp://dao.gsfc.nasa.gov/pub/papers/lyster/lagrangian.ps
Contents:
1. Introduction
2. The Lagranian filter
3. Interpolation and Trajectory Generation Algorithms
4. Parallelization and Performance Issues: including the HPCC Milestone and Speedup Curves
5. Behavior of the Lagrangian filter with the Baseline Data Set for September 1992
6. Summary and Conclusions

Download the source code: ftp://dao.gsfc.nasa.gov/pub/papers/lyster/lagn_euln_hpcc_r_1_0_0.tar.gz You can gunzip and untar this file and read lagn_euln/README for instructions on how to run it (specifically on an Origin 2000, but it is portable to other machines with the parallel MPI library).