Date and Location
The developer meeting is held on February 12-13, 2020 at IWR Room 1/414 in Heidelberg. We invite all DUNE developers and contributors to participate.
|Feb. 12||10:00 - 12:00||Meeting|
|12:00 - 13:00||Lunch|
|13:00 - 18:00||Meeting|
|Feb. 13||9:00 - 12:00||Meeting|
|12:00 - 13:00||Lunch|
|13:00 - 18:00||Meeting|
- Peter Bastian
- Markus Blatt
- Andreas Dedner (only Wednesday afternoon)
- Nils Dreier
- Christian Engwer
- Jorrit (Jö) Fahlke
- Christoph Grüninger (only Wednesday, Feb 12)
- René Heß
- Lasse Hinrichsen
- Dominic Kempf
- Ole Klein
- Robert Klöfkorn (only Wednesday afternoon)
- Santiago Ospina de los Rios
- Simon Praetorius
- Oliver Sander
- Linus Seelinger
- Michal Toth
Please use the edit function to add your topics or write an email to Markus.
Christoph: Vote on Simon becoming a Dune core developer
- In favour: 6
- Against: 0
- Abstain: 0
Christian: I suggest that we decide on a code-of-conduct. This should ensure that Dune is received as an open welcoming community. See for example the one of PETSc.
- Need particular proposal, need more information
Jö: What is the state of dune-gitlab?
Is currently staying in HD, Peter thinks there will be people in ~0.5 year to take care of a general IWR gitlab.
Until then, can we have more people with admin right for core? -> Rene can add people?
Jö: Seeking co-maintainers for mailing list servers (reduce bus factor) -> Markus
Christoph: Bump required compiler support from C++11 to C++14, i.e., GCC 4.9 to GCC 7 (from May 2017) and from ancient Clang 3.8 to Clang 5. Given Dune 2.8 will be released in 2021 or even later, this would enable us to use convenient C++ techniques and reduce the number of feature checks.
- Actually, support C++17 to the extent that GCC7 supports.
- Mingw: can live with GCC7
- Clang? 7 is in Debian buster
- Dumux using Clang 5
- Ubuntu Bionic Clang6
- Clusters not an issue anymore, usually have modules with newer compilers
Conclusion: C++17 to the extent supported by GCC7 and Clang5
- excludes replacement of class objects containing reference members (not supported in Clang5
- Christoph will make a list of more excluded features
- Who can update the CI system? Image building for CI system is broken currently. Simon will give it a try.
Christoph: Getting rid of dune-grid-howto. It’s in an embarrassing state. It could be replaced by Oliver’s book once this is published.
Possible alternative: “Recipes” by Linus.
Conclusion: dune-grid-howto won’t be part of the next release. Christoph will take care.
Christoph: I’d like to make a two-hour session to revise old merge request and issues. We should either decide what to do about them or close them. -> group work
Christoph: Reduce number of core modules, cf. mailing list thread
- RK: I think the number of core modules is ok. E.g. dune-istl should definitively stay separate from dune-grid. The one module I would like to see merged is dune-python into the various core modules as discussed during the last meeting.
- Christoph: I’d like to see dune-geometry and/or dune-istl merged to dune-common. Many aspects of istl and common are difficult to separate (defines of matrices, basic linear algebra). Dune-geometry is very specific / fine granular (geometry and quadrature rules) that it is not justified to be a module of its own. Including Python support sounds like a good idea.
- higher-order geometries might require localfunctions, and localfunctions requires geometry
- tooling to simplify releases?
- is there a limit in the number of core modules?
- merging or moving dune-uggrid? Remove splitting into dune-grid/dune-uggrid?
- Make dune-functions a core module? (not voted now)
dune-typetree -> dune-common (VOTE: yes) (need some cleanup)
Change of namespace in dune-typetree
VOTE on 2: 11, VOTE on 3: 6, Decision:
Markus: What ist the status of EXADUNE, what should we integrate back?
- Status: simd interfaces, vectorization of istl solvers, block-krylov methods, multi-threaded/GPU supported solvers (difficult to integrate), block-structured grids, higher-order methods, pipelined/communication-hiding krylov methods
- Integration: block-krylov methods (dune-istl), (partially) higher-order DG methods (dune-pdelab)
- backward iteration over grid elements (interface change?)
- multi-threaded istl solvers? (from kascade? from exadune? tbb-based?)
Linus: add a
module.hhheader in every module including everything from that module
- RK: too many dependencies -> lead to recompile everything
- don’t use this header in any other core module?
- benefits? beginners have to look for the headers to include
- cmake macro to generate header list?
- are there incompatible headers?
Simon: We have to update docker images to clang5 & test with c++17
Simon: Is there a place to check of user dune projects?
- OS: Modules page lack of structure (what is an extension?)
- OS: Include more descriptions to the modules (OS do it)
- Simon: Add more structure to the sections
- Have a web for each tag with the modules including such tag
Linus: examples should compile and run in the CI
- Very simple example witch is working and is documented in doxygen and that is also compiled and run in the CI
- There are enough tools in Doxygen to fully document specific parts of the code (e.g. snippet, link pages, show partial part of a file)
- See example: https://www.dune-project.org/doxygen/pdelab/master/group__pdelab__recipes.html
- This avoids that examples get outdated and give users the idea of how to use the software
- Use folder /doc/recepies/ so that is easy to link later in the webpage and add files to ´dune_test´
Linus: Is there interest for functionality to write tracing information in OTF-2 file-format?
- It records information the taks it is performing (entring and leaving a section in the code) and with the format one can get visualization of the load balance with VTE (there are other possibilities)
- With the OTF2 is a little bit complicated to instrument, so the proposal is to add a facility to trace easier code.
- Linus will send a merge request with the feature
Simon: Several links in the webpage are broken.
Not yet discussed:
- Jö: Cheat-sheet?
- Rober & Andreas: let’s talk about 3.0?
- Simon: Future plans?
Linus: Parallel communication using information about rank, some extension is needed in the data handles for the
- Nils and Robert are working on something already (although in a Private issue).
- Nils proposes a new interface (See https://gitlab.dune-project.org/core/dune-common/issues/174).
- GMSH reader:
- Jo referes to indexing problems with gmsh (See https://gitlab.dune-project.org/flyspray/FS/issues/1698)
- Simon: There is some work done to the gmsh format 4.0 (Parametrized geometry)
- GMSH 4 is not compatible with gmsh 2 format reader. Parallel information is the main change from it.
- The reader complains when it finds PhisicalEntities.
- Christian Use component readers for each section.
- The class VirtualFunction that fulfills the same purpose as std::function. We agree to deprecate it.
- viewThread proposal for the grid -> closed.
- GMSH reader:
- Christian: Can we get rid of FORTRAN? It is needed for LAPACK.
- Eigen & PETSc uses a LAPACK without fortran so we know it’s possible.
- Dominic: Discuss inclusion of (support for) a logging library
- The current implementation with streams does not attract much use
- Steffen left us a library: https://gitlab.dune-project.org/staging/dune-logging
- Other full fletched systems exist e.g. https://github.com/gabime/spdlog
- Christoph: Including dune-logging sounds good to me.
- Christian: I’m in favour of an external solution. We have to keep our maintenance effort in mind.
- Need for better logging support?
- Include libfmt (or something similar)?
- Needs more information! Comparison of logging libraries, more arguments for dune-logging, maintainability of dune-logging? hard or soft dependency? hierarchical logging? interface changes necessary to use loggers in dune classes?
- OS: The makers of CurvilinearGrid have relicensed their code, and the
license is now dune-compatible. Can we import their implementation of a
curvilinear geometry(!) into dune-geometry?
- Simon worked some with it
- They have a different numbering of langrange base nodes
- Simon is working on a extended geometry
- OS: ALUGrid implements an extended GridFactory that allows for
distributed grid setup. Can we make that an official interface?
- RK: Sounds like a good idea.
- Christoph: +1
- Here the interface changes we are talking about:
/** \brief insert a vertex into the coarse grid including the vertex's globally unique id * * \param[in] pos position of the vertex * \param[in] globalId globally unique id for vertex (e.g. globally unique vertex index) */ void insertVertex ( const VertexInputType &pos, const VertexId globalId ); // for ALUGrid it is also beneficial to have but this is not mandatory virtual void insertProcessBorder ( const std::vector< VertexId >& vertices );
- Additional information about rank sharing the vertices? - Decision: - Andreas: Submit proposal on GitLab to see the details
- OS: Interface clarification: when the grid changes, do
GridViewobjects remain valid? What about
- Example: YaspGrid: gridview is invalid after grid changes
- Carsten: See decision in dev-meeting 2009
- Simon: Langrange functions have unclear semantics, different implementations using inconsistent ordering of Lagrange nodes. Ordering should be documented.
- Simon: Factory and cache have very different interfaces, you can’t easily switch from one to the other.
- Nils: Parallel SolverFactory
- Polymorphic communication object
- User interface
- OS: I keep hearing from people that the RemoteIndex implementation
is showing its age. Can we organize a redesign/rewrite?
- Christian: The
ParallelIndexSetsare an aimplementation and not an interface. Nils and RK are currently working on a proposal for an interface, for which the
ParallelIndexSetscan be a particular implementation. My goal would be to have a coarse grained dynamic interface (in ISTL), which can then be fed with data from a
ParallelIndexSetor from a
Gridor something else…
- Christian: The
- Simon: Parallelization of dune-functions bases. Missing tools and utilities, e.g.
- Iteration over all DOFs on entities
- (global) IDs of DOFs
- Introduction of PartitionSet in several algorithms
- Something like Twist-Uitlities for uniquely identifying the order of DOFs on sub entities (also global unique)
- ParallelIndexSet for MultiIndices
- Maybe something like dune-grid:GlobalIndexSet but for the DOFs
- Properly wrapped local bases for DG discretization (association of local DOFs to the element and not to the sub-entities)
- OS: New ideas on how to support non-affine families of finite elements?
- Module: https://github.com/spraetor/dune-vtk
- OS: Simon has a module for reading and writing VTK stuff doing much more than the stuff integrated in dune-grid, including higher order functions and structured grid.
- This module can write specific vtk formats that are more compact. Paraview is much
faster when using these formats.
- Subsampling is still not implemented.
- Can write quadratic data (Not so easy to visualize in paraview)
- Structured and rectangular grids.
- PolyData not implemented (1D is in dune-grid because it was the only way possible to write 1D data)
- Christian: How does it compress data?
- Is stored in binary blocks compressed: collect - compress - write - come back and set the position of the data.
- Why don’t we use the VTK library?
- Is not very easy to use. Huge library. No smooth transitions between versions.
- Data collector: get grid info and its data, then decide whether to write continuous or discontinuous data. It is very flexible to be combined with other data collectors.
- Parallel grid format:
- If grid factory does have the global ID interface then it builds the grid on each processor
- otherwise it gets serialized in one processor and then load balance.
- Know Bug: When YaspGrid has periodic boundaries, it simply produces wrong results.
- In the long run the idea is to complete the missing features and to merge it into dune-grid.