A Method of Adaptive Coarsening for Compressing Numerical Data

Date:Friday, March 6, 2009
Time: 1:00pm - 2:00pm
Location: LBNL bldg. 50A, Room 5132

Speaker:
Scott B. Baden
 Department of Computer Science & Engineering
 University of California, San Diego

Abstract

    A challenge in large scale computing is how to compress simulation
    data without losing valuable information. While wavelet based
    methods are popular, they require that data be decompressed prior to
    analysis for example, when identifying time-dependent turbulent flow
    structures.

    We present adaptive coarsening, a lossy compression strategy based
    on adaptive subsampling that enables the compressed data product to
    be directly manipulated in memory. Like wavelet compression,
    adaptive coarsening is a multiresolution technique; however, it does
    not represent data progressively.

    We demonstrate compression factors of up to 8 in turbulent flow
    simulations in three dimensions. Our compression strategy produces a
    non-progressive multiresolution representation, subdividing the
    dataset into fixed sized regions and compressing each region
    independently.

Host of Seminar:

    Kathy Yelick