DM Basics#

The previous chapters have focused on the core numerical solvers in PETSc. However, numerical solvers without efficient ways (in both human and machine time) of connecting the solvers to the mathematical models and discretizations that people wish to build their simulations on, will not get widely used. Thus PETSc provides a set of abstractions represented by the DM object to provide a powerful, comprehensive mechanism for translating the problem specification of a model and its discretization to the language and API of solvers.

Some of the model classes DM currently supports are PDEs on structured and staggered grids with finite difference methods (DMDA and DMSTAGDMSTAG: Staggered, Structured Grid), PDEs on unstructured grids with finite element and finite volume methods (DMPLEXDMPlex: Unstructured Grids), PDEs on quad and octree-grids (DMFOREST), models on networks (graphs) such as the power grid or river networks (DMNETWORKNetworks), and particle-in-cell simulations (DMSWARM).

In previous chapters, we have demonstrated some simple usage of DM to provide the input for the solvers. In this chapter, and those that follow, we will dive deep into the capabilities of DM.

It is possible to create a DM with

DM dm;
DMCreate(MPI_Comm comm, DM *dm);
DMSetType(DM dm, DMType type);

but more commonly, a DM is created with a type-specific constructor; the construction process for each type of DM is discussed in the sections on each DMType. This chapter focuses on commonalities between all the DM so we assume the DM already exists and we wish to work with it.

As discussed earlier, a DM can construct vectors and matrices appropriate for a model and discretization and provide the mapping between the global and local vector representations.

The matrices produced may support MatSetValuesLocal() allowing one to work with the local numbering on each MPI rank. For DMDA one can also use MatSetValuesStencil() and for DMSTAG with DMStagMatSetValuesStencil().

A given DM can be refined for certain DMTypes with DMRefine() or coarsened with DMCoarsen(). Mappings between DMs may be obtained with routines such as DMCreateInterpolation(), DMCreateRestriction() and DMCreateInjection().

One attaches a DM to a solver object with

KSPSetDM(KSP ksp,DM dm);
SNESSetDM(SNES snes,DM dm);
TSSetDM(TS ts,DM dm);

Once the DM is attached, the solver can utilize it to create and process much of the data that the solver needs to set up and implement its solve. For example, with PCMG simply providing a DM can allow it to create all the data structures needed to run geometric multigrid on your problem.

SNES Tutorial ex19 demonstrates how this may be done with DMDA.