Advertisement

Parallelization of ILU decomposition for elliptic boundary value problem of the PDE on AP3000

  • Kentaro Moriya
  • Takashi Nodera
VII Poster Session Papers
Part of the Lecture Notes in Computer Science book series (LNCS, volume 1615)

Abstract

ILU (or Incomplete LU) decomposition is one of the most popular preconditioners for large and sparse linear systems of equations. However, it is difficult to implement the ILU preconditioner on distributed memory parallel computers, because the process consists of forward and backward substitution. The block divided method is one of the algorithms that can parallelize the ILU preconditioner for the linear system obtained by applying the finite difference method to discretize the elliptic boundary value problem of the PDE (or partial differential equation). However, on a distributed memory parallel computer, since the communication overhead is significantly large, the ILU preconditioner does not perform well. We propose an algorithm that decreases the communication overhead on the block divided method and determines the appropriate band-size. Based on our approach, the BiCGStab(ℓ) method with the ILU preconditioner is implemented on the distributed memory parallel computer, Fujitsu AP3000. We also analyze the performance of parallelism in the operation of the ILU preconditioner through numerical results.

Preview

Unable to display preview. Download preview PDF.

Unable to display preview. Download preview PDF.

References

  1. 1.
    Schönauer, W.: Scientific Computing on Vector Computers, North Holland (1987).Google Scholar
  2. 2.
    Wolfe, W.: More Iteration Space Tiling, Supercomputing'89, pp. 655–664 (1989).Google Scholar
  3. 3.
    Bastian, P. and Horton, G.: Parallelization of Robust Multigrid Methods: ILU Factorization and Frequency Decomposition Method, SIAM J. Sci. Stat. Comput., Vol. 12, No. 6, pp. 1457–1470 (1991).CrossRefMathSciNetMATHGoogle Scholar
  4. 4.
    Joubert, W.: Lanczos Methods for the Solution of Nonsymmetric Systems of Linear Equations, SIAM J. Matrix. Anal. Appl., Vol. 13, No. 3, pp. 928–943 (1992).MathSciNetCrossRefGoogle Scholar
  5. 5.
    Ramanujan, J. and Sadayappan, P.: Tiling Multidimensional Iteration Space for Multicomputers, J. of Parallel and Distri. Comp. Vol. 16, pp. 108–120 (1992).CrossRefGoogle Scholar
  6. 6.
    Sleijpen, G. L. G. and Fokkema, D. R.: BiCGStab(ℓ) for Linear Equations Involving Unsymmetric Matrices with Complex Spectrum, ETNA, Vol. 1, pp. 11–32 (1993).MathSciNetMATHGoogle Scholar
  7. 7.
    Bruaset, A. M.: A survey of Preconditioned Iterative Methods, Pitman Research Note in Mathematics, No. 32, Longman Scientific & Technical, U. K. (1995).MATHGoogle Scholar
  8. 8.
    Nodera, T. and Noguchi, Y.: Effectiveness of BiCGStab(ℓ) Method on AP1000, Transaction of Information, Processing Society of Japan (in Japanese), Vol. 28, No. 11, pp. 2089–2101 (1997).Google Scholar
  9. 9.
    Nodera, T. and Noguchi, Y.: A Note on the BiCGStab(ℓ) Method on AP1000, IMACS Series in Comp. and Applied Math., Vol. 4, pp. 53–58 (1998).Google Scholar
  10. 10.
    Nodera, T. and Tsuno, N.: The Parallelization of the Incomplete LU Factorization on AP1000, Lecture Note in Computer Science, Springer-Verlarg, Vol. 1470, pp. 788–792 (1998).Google Scholar
  11. 11.
    Vuik, K. and Van Nooyen, R. R. P.: A Parallel ILU-Preconditioner, IMACS Series in Comp. and Applied Math., Vol. 4, pp. 399–405 (1998).Google Scholar
  12. 12.
    Fujitsu Lab: AP3000: Products: Hardware, http://www.fujitsu.co.jp/hypertext/Products/Info_process/hpc/ap3000/products/hw.htm.Google Scholar

Copyright information

© Springer-Verlag Berlin Heidelberg 1999

Authors and Affiliations

  • Kentaro Moriya
    • 1
  • Takashi Nodera
    • 1
  1. 1.Keio UniversityYokohamaJapan

Personalised recommendations