Compatibility of scalapack with the discrete wavelet transform

Abstract

Parallelization of the Discrete Wavelet Transform (DWT) oriented to linear algebra applications, specially the solution of large linear systems, is studied in this paper. We propose that for parallel applications using linear algebra and wavelets, it can be advantageous to use directly the two-dimensional block cyclic distribution (2DBC), used in ScaLAPACK [1]. Although the parallel computation of the DWT may be faster with other data distributions, the efficiency of the 2DBC distribution for some linear algebra applications can make it more appropriate. We have tested this approach implementing a preconditioner for iterative solution of linear systems, which uses wavelets and was proposed in [2], The results using the 2DBC Distribution are better than those obtained with a popular parallel data distribution for computing wavelets, the "Communication-Efficient" scheme proposed in [3]. © Springer-Verlag Berlin Heidelberg 2007.