Atmospheric data assimilation on distributed-memory parallel supercomputers

Abstract

Atmospheric data such as temperature, moisture, winds, etc., collected by satellites and direct measuements from upper-air instruments, ground observation stations provide only partial information about the atmosphere. They are assimilated to numerical forecasts to provide a coherent, evolving state of the global atmosphere. The data analysis system, the Physical-space Statistical Analysis System (PSAS) developed at the Data Assimilation Office at NASA's Goddard Space Flight Center, requires computing resources far beyond the capabilities of even the state-of-the-art vector supercomputers. Here we describe an efficient and scalable implementation of the PSAS on distributed-memory massively parallel supercomputers such as Intel Paragon and Cray T3E; the implementations achieves superb performance as demonstrated by detailed performance analysis of systematic runs on up to 512 processors on Paragon, T3D and T3E. Consequently, the solution time is reduced to 24.6 seconds on 512-PE T3E from 5 hours on a single head of Cray C90 for a real problem of 80,000 observations, a 740-fold reduction of turn-around time.KeywordsData AssimilationMatrix BlockGrid RegionInnovation MatrixRequire Computing ResourceThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.