Parallel ISTL

This changes the linear algebra interface such that parallel linear algebra backends can be used and provides the implementation of the interface for parallel ISTL solvers.

Adds:

• Support for parallel ISTL solvers using the current solver interface
• Runtime switch for overlap type of the ISTL solver used
• solver->category: [default, sequential, nonoverlapping, overlapping] initfile parameter
• Comm object for the linear algebra interface that is to be used as a container for all required information for running the backend's solvers in parallel
• Implementation of the Comm object for ISTL

Changes:

• Linear algebra classes now use a Traits class as template parameter

TODO:

• Check if Grid uses overlap and use [Non]OverlappingSchwarzOperator
• Provide an efficient method for computing the parallel index set
• Resolve all TODO(FM) markers
• Make an alternative implementation for nonoverlapping Schwarz
• Handle the corner case overlapSize + ghostSize == 0
• Add an example or test file
• Update assembler element loops to use the proper partition set

Relates to #4 (closed).

added 1 deleted label

changed the description

added 1 commit

• 6285e6d5 - Added Comm template argument to LinAlg classes

Compare with previous version

added 1 commit

• 72ea2fb8 - Replaced Mat,VecX,VecB by Traits template parameter

Compare with previous version

added 1 commit

• 0ee4b343 - Replaced enable_if for DefaultCreators by template specialization

Compare with previous version

changed the description

added 1 commit

• 688b9563 - Simplified DefaultCreators template specializations

Compare with previous version

added 1 commit

• a573a982 - Simplified UmfpackRunner template specialization

Compare with previous version

I was considering splitting ISTLRunner into an #if HAVE_MPI part with the new things I changed in this MR and a sequential part that is basically what we had before. Then it will be easier to add sequential-only solvers later on. What are your thoughts on this?

added 1 commit

• 899b42a7 - Changed RunnerInterface to take an additional Comm argument; Added Comm...

Compare with previous version

That pipeline failure is due to a bug in dune/istl/owneroverlapcopy.hh, I'll fix that with a wrapper for the index pair later.

added 1 commit

• 5c2093af - Added workaround for comm index issue; Fixed problem with lack of copy elision for older compilers

Compare with previous version

marked the checklist item Check if Grid uses overlap and use [Non]OverlappingSchwarzOperator as completed

added 1 commit

• a5a50014 - Changed NoComm dummy to work with new interface

Compare with previous version

A general comment: Try to return std::unique_ptr in create() functions and name the corresponding classes Creators, not Factories.

added 1 commit

• ee683a4d - Moved typedefs to Traits class; Changed creator return type to unique_ptr

Compare with previous version