[Next]:  Three-dimensional Delaunay mesh generation  
 [Up]:  Projects  
 [Previous]:  Algorithms for the solution of the  
 [Contents]   [Index] 



Optimization of the sparse direct solver PARDISO  for use in the Intel Math Kernel Library

Collaborator: K. Gärtner

Cooperation with: O. Schenk (Universität Basel, IFI, Switzerland), B. Greer (Intel Corporation, Hillsboro, Oregon, USA)

Description: After passing the evaluation process at Intel Hillsboro, the incorporation of the solver into the Intel Math Kernel Library (MKL) was stated as project goal. The interplay of PARDISO  [1, 2, 3] with the Intel compilers and other Math Kernel Library routines (especially BLAS3 routines) was further improved. Intel took care of the questions arising from the programming interface conditions of the MKL and the problems related to redistributing the libraries for different processors and operating systems.

A MKL user interface introducing an additional abstraction layer is provided and, for experienced users, the original one is maintained, too. Because PARDISO  involves large sets of matrices with very different sizes, it will be used as one of the performance measurement tools at Intel. Future MKL development will particularly pay attention to the behavior in a parallel environment and will assure that especially the DGEMM implementation works as well as possible for small and large matrices.

Just now in January 2004 Intel announced the $ \beta$ release of PARDISO  within MKL

(see http://www.intel.com/software/products/mkl/beta/features.htm).

The distribution via the vendor library will benefit internal simulation needs and provide users outside academia with a stable code basis.


Fig. 1: Elimination tree after permutation for four processors
\ProjektEPSbildNocap{11cm}{a_t12.tree_cmd_super4.eps}

References:

  1. O. SCHENK, K. GÄRTNER, W. FICHTNER, Efficient sparse LU factorization with left-right looking strategy on shared memory multiprocessors, BIT, 1 (2000), pp. 158-176.
  2. O. SCHENK, Scalable Parallel Sparse LU Factorization Methods on Shared Memory Multiprocessors, ETH Zürich, Ph.D. Thesis 2000.
  3. O. SCHENK, K. GÄRTNER, Solving unsymmetric sparse systems of linear equations with PARDISO,  Future Generation Computer Systems, in press.



 [Next]:  Three-dimensional Delaunay mesh generation  
 [Up]:  Projects  
 [Previous]:  Algorithms for the solution of the  
 [Contents]   [Index] 

LaTeX typesetting by I. Bremer
2004-08-13