Meeting April 2006

Participants

Alessandro Veneziani (AV), Miguel Fernandez (MF), Christoph Winkelmann (CW), Christophe Prud'homme (CP), Martin Prosi (MP), Gonçalo Pena (GP)

Transition to new math kernel

CP presents the state of the new kernel and current issues.

Tests

AV presents results by Christian Vergara and Tiziano Passerini comparing the solutions of PC and IP solver on a flux imposed Navier-Stokes problem for two different bifurcation geometries. PC takes about 10 times longer while the solutions differ by 10%, and the global mass balance is better satisfied by the solution from PC. See full presentation here (PDF).

CW points out that we are comparing two solvers which have different strategies for solving the linear system (mixed approach vs. monolithic) and for spacial discretization (stable vs. stabilized approach). The comparison does not tell us which of these aspects is responsible for which property of the solver. For more insight, we would need more orthogonalized solvers (see below)

The workgroup decides to carry out more systematic comparisons of the different solvers. Until May 15th, three solvers (PC, IP, SUPG) are going to be compared on three different problems: the Turek cylinder benchmark (by CW), the Womersley flow (by MF) and the physiological problem already investigated (by C. Vergara and T. Passerini). The Womersley test should become the test in the testsuite replacing several current ones. The two other test cases are to be put in playground.

Re-Design and Orthogonalization

CW shortly presents his attempts to rewrite the stabilized solvers in an orthogonal way. The aim is to minimize code duplication, reducing the maintenance efforts and increasing the modularity of the code, which allows for testing new combinations. The components that should be independent are:

• finite element space
• time discretization
• type of equation (NS/Oseen)
• stabilization
• linear solver

The current state can be found in playground, folder winkelma.

The workgroup decides that MF and CW work on the sketch on Tuesday, providing solutions for moving meshes and different stabilizations. This objective has been reached, and the results can be found in playground, folder winkelma. Another question is the merge of stable and stabilized solvers, which turned out to require substantial modifications in the matrix classes of LifeV. As they are known to be flawed and will be replaced with the new kernel anyway, it is highly questionable whether a full decoupling is worth the effort in this domain.