From: bangerth Date: Fri, 7 Mar 2008 14:27:21 +0000 (+0000) Subject: Document more results obtained with the various renumbering strategies. X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=6cb0f28d303328459525a8d68681900cd36b3927;p=dealii-svn.git Document more results obtained with the various renumbering strategies. git-svn-id: https://svn.dealii.org/trunk@15868 0785d39b-7218-0410-832d-ea1e28bc413d --- diff --git a/deal.II/deal.II/include/dofs/dof_renumbering.h b/deal.II/deal.II/include/dofs/dof_renumbering.h index 275115c012..677fbbef27 100644 --- a/deal.II/deal.II/include/dofs/dof_renumbering.h +++ b/deal.II/deal.II/include/dofs/dof_renumbering.h @@ -203,6 +203,36 @@ DEAL_II_NAMESPACE_OPEN * *

A comparison of reordering strategies

* + * As a benchmark of comparison, let us consider what the different + * sparsity patterns produced by the various algorithms when using the + * $Q_2^d\times Q_1$ element combination typically employed in the + * discretization of Stokes equations, when used on the mesh obtained + * in @ref step_22 "step-22" after one adaptive mesh refinement in + * 3d. The space dimension together with the coupled finite element + * leads to a rather dense system matrix with, on average around 180 + * nonzero entries per row. After applying each of the reordering + * strategies shown below, the degrees of freedom are also sorted + * using DoFRenumbering::component_wise into velocity and pressure + * groups; this produces the $2\times 2$ block structure seen below + * with the large velocity-velocity block at top left, small + * pressure-pressure block at bottom right, and coupling blocks at top + * right and bottom left. + * + * The goal of reordering strategies is to improve the + * preconditioner. In @ref step_22 "step-22" we use a SparseILU to + * preconditioner for the velocity-velocity block at the top left. The + * quality of the preconditioner can then be measured by the number of + * CG iterations required to solve a linear system with this + * block. For some of the reordering strategies below we record this + * number for adaptive refinement cycle 3, with 93176 degrees of + * freedom; because we solve several linear systems with the same + * matrix in the Schur complement, the average number of iterations is + * reported. The lower the number the better the preconditioner and + * consequently the better the renumbering of degrees of freedom is + * suited for this task. We also state the run-time of the program, in + * part determined by the number of iterations needed, for the first 4 + * cycles on one of our machines. + * * * * + * + * * * * * * + * * * * * + * + * * * * * *
@@ -214,43 +244,87 @@ DEAL_II_NAMESPACE_OPEN * * @image html "reorder_sparsity_step_31_deal_cmk.png" *
- * @image html "reorder_sparsity_step_31_boost_cmk.png" + * Enumeration as produced by deal.II's DoFHandler::distribute_dofs function + * and no further reordering apart from the component-wise one. + * + * With this renumbering, we needed an average of 92.2 iterations for the + * testcase outlined above, and a runtime of 7min53s. * - * @image html "reorder_sparsity_step_31_boost_king.png" + * Random enumeration as produced by applying DoFRenumbering::random + * after calling DoFHandler::distribute_dofs. This enumeration produces + * nonzero entries in matrices pretty much everywhere, appearing here as + * an entirely unstructured matrix. + * + * With this renumbering, we needed an average of 71 iterations for the + * testcase outlined above, and a runtime of 10min55s. The longer runtime + * despite less iterations compared to the default ordering may be due to + * the fact that computing and applying the ILU requires us to jump back + * and forth all through memory due to the lack of localization of + * matrix entries around the diagonal; this then leads to many cache + * misses and consequently bad timings. * - * @image html "reorder_sparsity_step_31_boost_md.png" + * Cuthill-McKee enumeration as produced by calling the deal.II implementation + * of the algorithm provided by DoFRenumbering::Cuthill_McKee + * after DoFHandler::distribute_dofs. + * + * With this renumbering, we needed an average of 57.3 iterations for the + * testcase outlined above, and a runtime of 6min10s. + *
- * Enumeration as produced by deal.II's DoFHandler::distribute_dofs function + * @image html "reorder_sparsity_step_31_boost_cmk.png" * - * Random enumeration as produced by applying DoFRenumbering::random - * after calling DoFHandler::distribute_dofs. + * @image html "reorder_sparsity_step_31_boost_king.png" * - * Cuthill-McKee enumeration as produced by calling the deal.II implementation - * of the algorithm provided by DoFRenumbering::Cuthill_McKee - * after DoFHandler::distribute_dofs. + * @image html "reorder_sparsity_step_31_boost_md.png" *
* Cuthill-McKee enumeration as produced by calling the BOOST implementation * of the algorithm provided by DoFRenumbering::boost::Cuthill_McKee * after DoFHandler::distribute_dofs. + * + * With this renumbering, we needed an average of 51.7 iterations for the + * testcase outlined above, and a runtime of 5min52s. * * King enumeration as produced by calling the BOOST implementation * of the algorithm provided by DoFRenumbering::boost::king_ordering - * after DoFHandler::distribute_dofs. + * after DoFHandler::distribute_dofs. The sparsity pattern appears + * denser than with BOOST's Cuthill-McKee algorithm; however, this is + * only an illusion: the number of nonzero entries is the same, they are + * simply not as well clustered. + * + * With this renumbering, we needed an average of 51.0 iterations for the + * testcase outlined above, and a runtime of 5min03s. Although the number + * of iterations is only slightly less than with BOOST's Cuthill-McKee + * implementation, runtime is significantly less. This, again, may be due + * to cache effects. * * Minimum degree enumeration as produced by calling the BOOST implementation * of the algorithm provided by DoFRenumbering::boost::minimum_degree - * after DoFHandler::distribute_dofs. + * after DoFHandler::distribute_dofs. The minimum degree algorithm does not + * attempt to minimize the bandwidth of a matrix but to minimize the amount + * of fill-in a LU decomposition would produce, i.e. the number of places in + * the matrix that would be occupied by elements of an LU decompisition that + * are not already occupied by elements of the original matrix. The resulting + * sparsity pattern obviously has an entirely different structure than the + * ones produced by algorithms trying to minimize the bandwidth. + * + * With this renumbering, we needed an average of 58.9 iterations for the + * testcase outlined above, and a runtime of 6min11s. *