This presentation discusses the NASA data assimilation project at the Data Assimilation Office at the NASA/Goddard Space Flight Center. The goal is to produce accurate gridded datasets of atmospheric fields by assimilating a range of observations along with physically consistent model forecasts. This work produces datasets that are used by the climate research community. The data come from conventional sources that are used for weather forecasts (e.g., radiosondes, earth-surface measurements, and satellite temperature retrievals), as well as new sources such as satellites that will be launched under the Mission To Planet Earth Enterprise. An end-to-end Goddard Earth Observing System (GEOS) Data Assimilation System (DAS) currently supports stratospheric flight missions and reanalysis projects for NASA. The current Core of this system (Model, and Analysis) is a multitasking algorithm that runs on Cray J90 and C90 computers at Goddard and NASA Ames Research Center. Future Core computing will be carried out at Ames, with a new production system scheduled to be ready for the EOS AM-1 satellite launch in June of 1998. The DAO has acquired SGI Origin 2000 computers, with an aggregate of 160 processors in place at Ames, and more planned for the future. The DAO is currently updating the control scripts and programs, and implementing a modular Fortran 90 Core system. During 1998 the Core system will be migrated to distributed-memory software using the Message Passing Interface. Part of this work is being carried out under the NASA High Performance Computing and Communications Earth and Space Sciences program. The algorithmic and performance issues involved in Core system are the main subject of this presentation.

An MPI implementation of the Basic Linear Algebra Communication Subprograms (BLACS), an underlying layer of the ScaLAPACK library is presented. Use is made of a wide spectrum of functionality available in MPI to realize BLACS as succinctly as possible, thus making the implementation concise, but still yielding good performance. Some of the implementation details are discussed and the benchmark results for the ScaLAPACK LU factorization on several parallel architectures with different MPI libraries are presented. A performance comparison with other existing BLACS implementations is made and some conclusions are drawn from the results.