Role of precision in meteorological computing: a study using the NMITLI Varsha GCM

Venkatesh, TN and Sinha, UN (2007) Role of precision in meteorological computing: a study using the NMITLI Varsha GCM. In: "Use of high performance computing in meteorology" Edited by George Mozdzynski. World Scientific.

Full text available as:
[img] PDF
Download (619Kb)

    Abstract

    An issue in large scale computing of unsteady phenomena is the role played by round-off errors. These errors can accumulate in long integrations leading to major differences in the computed quantities such as rainfall. Computations with the same code at different precisions (single, double and arbitrary precision) have been made to assess the role of round-off. For the study, Varsha2C and Varsha2C-MP, C and multi-precision versions of the spectral, hydrostatic Varsha GCM have been used. Varsha2C-MP uses the ARPREC library for computing using an arbitrary number of digits. We find that even for sequential runs, round-off causes solutions to differ beyond 12 days when two different processors are used. With the same processors, the results of parallel runs made using a different number of processor s differ significantly after 10 days, due to the round-off errors in the global summation operations. Use of double precision and changing the communication strategy reduces the differences by about 10 %. With the use of multi-precision, repeatable results, both across different processor types and different number of processors, can be obtained for up to 30 days of integration, provided 64 or more digits are used for the computing. For the Lorenz system of equations, however, precision has no significant effect on the results. Calculations using up to 2048 digits show that given the same initial perturbation, the time at which two solutions diverge (by order 1) does not depend on the precision. These results suggest that, for long integrations of meteorological models, far more careful computing is required than is generally appreciated.

    Item Type: Book / Book Chapter
    Subjects: GEOSCIENCES > Meteorology and Climatology
    MATHEMATICAL AND COMPUTER SCIENCES > Mathematical and Computer Scienes(General)
    Division/Department: Flosolver Division, Flosolver Division
    Depositing User: TN Venkatesh
    Date Deposited: 24 Jan 2012 10:04
    Last Modified: 24 Jan 2012 10:04
    URI: http://nal-ir.nal.res.in/id/eprint/10256

    Actions (login required)

    View Item