next up previous contents index
Next: 6. Linear-Algebraic System Interface Up: 5. Finite Element Interface Previous: 5. Finite Element Interface   Contents   Index

Subsections

5.1 Iterative methods and preconditioners available

5.1.1 Iterative methods

  1. Krylov solvers (conjugate gradient, GMRES, TFQMR, BiCGSTAB)
  2. BoomerAMG (a parallel algebraic multigrid solver)
  3. SuperLU direct solver (sequential)
  4. SuperLU direct solver with iterative refinement (sequential)

5.1.2 Preconditioners

  1. diagonal
  2. parallel incomplete LU with threshold (PILUT)
  3. another parallel incomplete LU (Euclid)
  4. parallel algebraic multigrid (BoomerAMG)
  5. parallel sparse approximate inverse (ParaSails)


Table 5.1: Parameters.
Parameter Name Parameter Values
solver cg, gmres (default), bicgstab, tfqmr, boomeramg, superlu, superlux
preconditioner diagonal (default), pilut, parasails, boomeramg, euclid
gmresDim an integer specifying the value of m in restarted GMRES(m). The default value is 50.
maxIterations an integer specifying the maximum number of iterations permitted for CG or GMRES. The default value is 1000.
tolerance a floating point number specifying the termination criterion for CG or GMRES. The default value is 1.0E-10.
pilutFillin an integer specifying the maximum number of nonzeros kept in the formation of incomplete L and U. If this is not called, a value will be selected based on the sparsity of the matrix.
pilutDropTol a floating point number specifying the threshold to drop small entries in L and U. The default value is 0.0.
euclidNlevels a non-negative integer specifying the desired sparsity of the incomplete factors. The default value is 0.
euclidThreshold a floating point number specifying the threshold used to sparsify the incomplete factors. The default value is 0.0.
superluOrdering natural (default) or mmd (minimum degree ordering). This ordering is used to minimize the number of nonzeros generated in the LU decomposition. The default is natural ordering.
superluScale y (yes; perform row and column scalings before decomposition) or n (no; default).
amgCoarsenType falgout, ruge, or default (CLJP) coarsening for BoomerAMG.
amgNumSweeps an integer specifying the number of pre- and post-smoothing at each level of BoomerAMG. The default is one pre- and one post-smoothings.
amgRelaxType jacobi (Damped Jacobi), gs-slow (sequential Gauss-Seidel), gs-fast (Gauss-Seidel on interior nodes), hybrid, or direct. The default is hybrid.
amgRelaxWeight a floating point number between 0 and 1 specifying the damping factor for BoomerAMG's damped Jacobi smoother. The default value is 0.5.
amgStrongThreshold a floating point number between 0 and 1 specifying the threshold used to determine strong coupling in BoomerAMG's coasening. The default value is 0.25.
parasailsThreshold a floating point number between 0 and 1 specifying the threshold used to prune small entries in setting up the sparse approximate inverse. The default value is 0.0.
parasailsNlevels an integer larger than 0 specifying the desired sparsity of the approximate inverse. The default value is 1.
parasailsFilter a floating point number between 0 and 1 defining the threshold used to prune small entries in A. The default is 0.0.
parasailsLoadbal a floating point number between 0 and 1 specifying how load balancing has to be done. The default is 0.0.
parasailsSymmetric set ParaSails to take A as symmetric.
parasailsUnSymmetric set ParaSails to take A as nonsymmetric (default).
parasailsReuse set ParaSails to reuse the sparsity pattern of A.



next up previous contents index
Next: 6. Linear-Algebraic System Interface Up: 5. Finite Element Interface Previous: 5. Finite Element Interface   Contents   Index
Tom Treadway 2003-10-22