GitLab

#1443 Test LoadPrepartitionedGodOfDof: the reason of the failure with Quadrangles elements is the mesh which was too small for being spread over so many processors. A ticket already exists (unfortunately can't be referenced currently due to Redmine being currently unavailable) to handle more properly the issue of meshes too small.

#1443 LoadPrepartitionedGodOfDof test: add another case with a mesh with quadrangles... which does not pass yet.

- A vectorial unknown
- A second FEltSpace in which both unknown are in the same NumberingSubset.
- Volume dofs (through P0)

#1443 Fix the test about program-wise dofs per numbering subset. Now there are some clean-ups to do (print to remove, comments to add, etc...).

- Now Dof holds a weak_ptr toward its Node, which itself gets one toward its NodeBearer.
- Whether a Dof belongs to a NumberingSubset is now handled by a test upon its Node.
- Dof program-wise index by numbering subset are now mandatory to fill processor-wise index by numbering subset - this is to avoid invalid state of the object, which is quite complicated to build (part of it before partitioning, and part
of it after).

#1443 WIP: the test still fails due to a chicken and egg problem: I need to isolate only Dofs related to a numbering subset to renumber them, but to know their NumberingSubset I need first their index... To break that I will change the weak dependancy: currently a Dof has a weak pointer to the NodeBearer, shortcircuiting the Node in the process... too bad the Node is the one to get the intel I need! I will therefore make the weak pointer toward the node, which itself will point toward the NodeBearer.

#1443 As I tried to go further (and reorganised a bit the GodOfDof code to avoid too much work in the future for the DRY refactoring...) I stumbled on an issue: we also need to set properly the program-wise dof indexes per numbering subset. I have added the test bit not the resolution.

#0 Doxygen: change order of comment to get rid of of an unfounder Doxygen warning for basic documentation.

#1443 Two steps parallelism: make the number of dofs match between reconstructed and original runs (for program-wise ones this entails saving the data, as it can't be recomputed from partial data - or not in an efficient way in the very least.

#1443 Move stuff into anonymous namespace to make the actual method implementation more straightforward.

#1443 Fix a bug in two steps parallelism: one index could be by mistake attributed to processor-wise rather than to ghosted.

#1443 NodeBearer after prepartition: fill propertly the processor_attribute (test has been updated along the way).

#1443 Complete the tests for NodeBearer in LoadPrepartitionedGodOfDof - and fix the bugs regarding the indexing of ghosted node bearers (previously they were numbered the same way as the processor-wise one, but it doesn't make any sense... Now their original index is also given in the prepartitioned data and reassigned properly during the renumbering.

#1443 Two step parallelisms in progress for GodOfDof: Node Bearer are seemingly properly reconstructed (test should be enriched to be absolutely sure of this).

#1443 Write in dedicated files informations that should allow proper reassignment of node bearers. It clearly won't be very efficient - the goal for now is to reach workable 2 step parallelism.

#1522 Mesh constructor from prepartitioned data: remove a call to ShrinkToProcessorWise which was actually harmful, and replace it by a more appropriate method.

Lua file for the LoadFromPrepartitionedMesh has been modified to feature P2 finite elements (as it was this that triggered the bug discovery).

#1443 Introducing a new class to write god of dof related data for prepartitioning. I however have first to investigate an issue when P2 is used (this choice is anyway more relevant for this test than P1...)

#1443 Rename more accurately NodeBearer index; also add safety by initializing the data attribute in the class declaration.

#1443 Add temporary catch in the test to take into account the temporary GracefulExit exception introduced in two step parallelism.

#1443 Start writing the test about prepartitioned data and GodOfDof (mostly it partitiones the data, write off the reults and load the new GodOfDof at the moment without cheking its content - which is obviously incorrect as we habve not yet implemented the proper init method.

In the process, new method Create() has been added in GodOfDofManager, to be used only in the test (or internally to respect DRY principle)...

#1443 Add a new test for loading GodOfDof from prepartitioned data (it's an empty shell for now...).

#1519 Remove AnalyticalInternalVariableOperatorCheckResults, which was not using relevant expected results.