Depending on the Dune version, the boundary markers are present on
all processes (<= 2.6) or on the root process only (>= 2.7). Try to
handle this in a flexible way in the corresponding data handle:
Determine if the minimum size over all processes of the boundary
markers vector is zero. If yes, assume that the root process contains
all markers and broadcast them.

Currently, it isn't possible to refine a parallel UGGrid that has
been read by a GmshReader, see dune-grid issue #83. Therefore, don't
refine in the corresponding test.

Kilian Weishaupt authored 6 years ago and Timo Koch committed 6 years ago

* test failed on master with old 360 deg 3d ref solution

* fixes clang compiler error

Implement a grid manager class that creates grids and a grid data object. The new class
is no longer a singleton. It is instantiated by the user. It is possible to get a
shared_ptr to the grid data to pass it on to e.g. problem / spatial parameter classes.

* No longer sets the overlap automatically (overlap to be checked by the grid geometry)
* Does not depend on discretization method or TypeTag anymore

Joint work with Kilian Weishaupt <kilian.weishaupt@iws.uni-stuttgart.de>

Bernd Flemisch authored 6 years ago and Timo Koch committed 6 years ago

The distribution of a grid over several cores can be system-dependent,
rendering the direct comparison of resulting parallel vtk files
impossible. To enable a comparison, convert the parallel files into
sequential ones in the fuzzycompare Python script. With this approach,
it is possible to compare the parallel output against the sequential
one. Adapt the gridcreator tests correspondingly.

Concerning the converted vtk files, a few difficulties had to be
handled:
* The files usually have duplicated vertices at the process borders.
  Eliminate the duplicated entries.

* The files can have an empty `PointData` section yielding to problems
  since the number of children differs. Eliminate the empty section
  from the files.

* The attributes may be different for each node: the
  `NumberOfComponents` isn't stated explicity if it is equal to `1`.
  On the other hand, the minimum and maximum values are provided
  which is not the case for the original vtk output. Eliminate the
  check of the attributes if a conversion from pvd has happened.

Enhance the gridcreator tests using boundary markers with an
additional synchronization step for those markers in the parallel
regime.

Bernd Flemisch authored 6 years ago and Timo Koch committed 6 years ago

Make the tests more flexible by passing a bool to refine or not.
Don't refine in the case of the ALUGrid 3d test, since apparently
the boundary parametrization isn't handled correctly during
refinement. Adapt the reference solutions.

Bernd Flemisch authored 6 years ago and Timo Koch committed 6 years ago

Realize boundary marker load balancing for ALUGrid. To achieve this, adapt
the data handle such that also the boundary markers are balanced. Since
the data handle works on faces rather than intersections, transfer first
the vector that is given in terms of boundary segments to a vector holding
a parameter value for each mesh face. After balancing, transfer the data
back to a boundary segment vector.

Restructure the tests and add more tests. Both UG and ALU are now tested
for Gmsh 2/3d and DGF 2d. The ALU tests for Gmsh 3d still need to be
worked on.

This test should also work in parallel now and a parallel test is added