Craig's home page
UKy  Yale
Preprints The family
ML-DDDAS Group Free software Digital pictures
  My Current Class Old Classes
 MGNet IMA 2002 Graduate Student Summer Program
DDDAS LNCC-CCS Workshop
Minimus Journal Computing


Sandia National Laboratories ASCI Project

Fast Cache-Aware Algorithms and Their Implementations

    As part of two projects funded by the ASCI and CSRI programs of Sandia National Laboratories from 2001-2004, we have studied algorithms to solve partial differential equations on computers with hierarchical memory systems.  Such computers replicate small pieces of main memory in caches.  Very fast computers currently have deep memories, where there are are usually two or more caches between the central processing unit (CPU) and the main memory.

    Algorithms to use caches efficiently have been developed that are bitwise compatible with standard implementations.  By bitwise compatible, we mean that we compute exactly the same solution.  Mathematically this means that if the cache aware solution is uc and the standard solution is us, then || uc - us || = 0.  While the implementations of algorithms like multigrid or Gauss-Seidel are very different, the actual order of the computations are identical.

    The goal of the ASCI funded grant was to support a research assistant: Danny Thorne.  Danny completed his Ph.D. in December, 2003.  His major contributions are his dissertation and cache aware hierarchical multilevel solver:

  • D. T. Thorne, Multigrid with Cache Optimizations on Adpative Refinement Mesh Hierarchies, Ph.D. dissertation, University of Kentucky, Department of Computer Science, Lexington, KY, December, 2003.
        PostScript     PDF
     
  • Software:  November, 2003 snapshot
        gzipped tar file

    A number of papers have been produced as part of this project, including

  • C. C. Douglas, J. Hu, J. Ray, D. T. Thorne, and R. S. Tuminaro, Fast, adaptively refined computational elements in 3D, in Computational Science - ICCS 2002, P. M. A. Sllot, C. J. K. Tan, J. J. Dongarra, and A. G. Hoekstra (eds.), 3 (2002), pp. 774-783.
  • C. C. Douglas and D. T. Thorne, A note on cache memory methods for multigrid in three dimensions, Contemporary Mathematics, 306 (2002), pp. 167-177.
  • D. T. Thorne, Cache Aware Multigrid on AMR Hierarchies.  Appeared in the Virtual Proceedings of the Copper Mountain Conference on Multigrid Methods, MGNet Virtual Proceedings Series, Cos Cob, 2003, 10 pages.
  • C. C. Douglas, J. Hu, J. Ray, D. T. Thorne, and R. Tuminaro, Cache aware multigrid for variable coefficient elliptic problems on adaptive mesh refinement hierarchies, to appear in Numerical Linear Algrebra and Applications, 2004.
  • C. C. Douglas, J. Hu, J. Ray, R. S. Tuminaro, and D. T. Thorne, Cache aware multigrid on two dimensional adaptively refined structured meshes, submitted September, 2003.
  • C. C. Douglas, G. Haase, and J. Hu, The state of quasi-unstructured grid cache aware multigrid, submitted June, 2002.

Cheers,
Craig C. Douglas

Last modified: