March 2017 PDELab Meeting Minutes

Participants: Andreas Nüssing, Christian Engwer, Jorrit Fahlke, Oliver Sander, Steffen Müthing

Later Carsten also discussed a few things via phone

Decision Process

  • There is a lack of feedback, leading to frustration for contributors and less contributions
  • There a conflicting requirements: Lectures need a stable base for their codes, but this can block new developments
    • Do we want a “PDELabNT”? No, that would mean less testing for older functionality.
  • Do we want to require “PDELab extension proposals” for bigger changes, i.e. small writeups explaining what is to be done? This may be a good thing sometimes, but ultimately we trust the proposer to choose an appropriate format for the proposal (i.e. extension proposal, issue, or possibly even another repository)

Contributions

  • Every developer should take responsibility: we need to give more feedback, in particular on merge requests
  • Do we want to organize sprints? We need to find a date early, three months before is probably too late. Mid of September appear to be a promising date among the people present.
  • Do we want to do more virtual bug-squashing weeks? If so
    • we should do occasional video-conferences during the week. We used Jitsi during the Dune Developermeeting, which is at least semi-open. There were some problems with the audio though, maybe a microphone with integrated noise cancellation will help. A dedicated camera would also be helpful for such events.
    • we should have a chat system where everyone is present during the week
    • The list of bugs to work on should be present before the event.

Road-Map PDELab

This is stuff that can move to PDELab proper soonish, and does not require any dune-functions-related facilities.

  • [Jö] Adept local operator wrappers: jacobian_*() from alpha_*() via automatic differentiation. #87
  • [Marian] Marian has some code for matrix-free solvers for nonlinear problems. Documentation is currently lacking though and needs writing. #88
  • [Steffen] Entity sets (in the sense “set of entities of multiple codimensions plus IndexSet") can be moved to core Dune as special GridViews. #89
  • [Steffen] Entity sets can be removed from PDELab as soon as they are in core dune. They need to stay as a concept though(?). There is built-in MCMGMapper-functionality in the IndexSet which needs to be removed. #90
  • [Steffen] Deprecate BoundaryGridFunction #91

Requirements for a new (dune-functions-inspired) Interface

This is stuff that makes sense only in connection with dune-functions

  • [Andreas] GridOperator on subsets?
  • [Marian] Alternative onestep grid-operator, which passes the weights of the spatial and the temporal parts to the single local operator so functions need to be evaluated only once. There is code in EXADUNE, but that needs to be extended before general consumption. See #92. Possibly relates: #39
  • [Steffen] Kill LocalAssemblerCallSwitch. #93
  • Assembler
    • [Christi] Cut-Cell stuff is too special, the Cut-Cell module should do its own assembling. Probably by hooking into quadratureRule().
    • [Christi] What is needed for GridGlue, in particular entities with differing C++ types?
    • [Steffen] What about MultiDomain? Spaces should no longer be restricted to subdomains, only GridFunctionSpaces.
  • [Steffen] Clean up gridfunctionspaceutilities.hh. #94.

List of Current Problems or Works-in-Progress

  • The interface of local operators needs revamping #42
  • Documentation, in particular
    • Constraints assembler
    • ConstraintsContainer should be considered a (very sparse) matrix. It should be documented as such. See also cc-interface.
    • There is a lack of conceptual documentation for developers, describing systems of interacting classes. In particular: the orderings (see also #71, #55).
  • There is some code duplication for the sole purpose of providing Doxygen documentation
  • The intorder parameter in adapt()
  • Defaulting to NumericalJacobian, see #45. This should not be tolerated in the future, since it leads to surprising behaviour. a) PDELab-provided local operators that use NumericalJacobian are considered buggy, see #85 b) NumericalJacobian should no longer be advertised, e.g. in courses c) Users that still want to use a numerical jacobian should wrap their local operators with add_numerical_jacobian(lop) or similar, see #86 d) Provide similar wrapper constructors for automatic differentiation.
  • The constraints containers need an interface for external consumption. This would enhance discovarability and can aid in debugging. It would also allow to use the information they contain in ways not forseen inside PDELab. See also cc-as-matrix.
  • Constraints need to improve composability: sometimes you need the unconstrained matrix in addition to the contrained one, e.g. for error estimators. A constrained function space should be considered a different function space, just with the same base.
  • The ISTL-backends need to clean up their namespacing. The solverbackends are going to be revamped completely, which is why they have not moved to the same namespace as the vector/matrix backends. Oliver Sander is considering doing a merge request that aliases the current solver backends into the ISTL backend namespace for consistency of description in his book.

What parts of PDELab is Oliver going to use/touch in his Book

(Obviously, this is the current planning, so things might change)

  • dune-functions shim
  • local operators
  • Dirichlet boundary conditions, but not general constraints
  • backends
  • adaptivity
  • an outlook on how to deal with systems

Notes

  • [Andi Nüssing] The group in Münster (Engwer) has some documentation on orderings in its internal Wiki
    Note: this later turned out to be pretty useless. Maybe Sebastian Westerheide has some more useful documentation somewhere, we’ll have to ask him.
  • [Christi] Do we want to support other kinds of interpolation/projection in dune-localfunctions?
  • [-> dune-functions] IntersectionSets?

Phonecall with Carsten

On BoundaryGridFunction

  • fufem may have something like a “trace of a base”
  • fufem has “boundary patches”, which mark a part of a boundary. They are sets of boundary intersections an can be iterated over.
  • fufem uses separate functions for extracting Dirichlet and Neuman boundary condition values
  • fufem does not have something like a BoundaryGridFunction, it would use a volume function instead
  • dune-function’s GridFunction would probably work just fine when using an IntersectionSet in place of an EntitySet. Analytic GridFunctions though are probably using a GridView, they would need to be adapted.
  • A DiscreteGridFunction for dune-functions and intersections is difficult at the moment, because intersections don’t have indices in dune-functions.

Using dune-functions' Discrete Functions in PDELab

Do we need/how to do .setTime() in dune-functions?

  • Consensus: we do not want .setTime(...) syntax in dune-functions
  • Proposal by Carsten: extend the interface from Range(Domain) to Range(Domain, Args...), plus bind/currying. Think “parameterised function”. Provide only differentiation with regards to Domain, not Args...
  • Proposal by Christian: Require that derivatives are only obtained from unary functions. If you have an n-ary function you need to bind/curry the other arguments first
  • Proposal by Carsten: Special dune-functions bind(), which allows to bind arguments after Domain
  • Consensus: Keeping an n-ary interface also in the lower levels is easiest.
Creative Commons License   |  Legal Statements / Impressum  |  Hosted by TU Dresden  |  generated with Hugo v0.80.0 (Mar 28, 23:25, 2024)