Which Clawpack solver should I use?

Clawpack includes a number of related hyperbolic PDE solvers:

All of them are built on common algorithmic ideas, make use of the same set of Riemann solvers, and can be used with VisClaw for visualization. If you’re not sure which solver to use, here you will find the main differences between them.

Installation and user interface

The AMRClaw, GeoClaw, and Classic solvers are traditional Fortran-based packages and rely on Makefiles and environment variables. Problems are specified partially through Python scripts (setrun.py) and partially through custom Fortran code (to set initial conditions, for instance).

With PyClaw, problems are specified through Python script files, or interactively (e.g., in IPython). Typically, the user does not need to write any Fortran code (though custom routines can be written in Fortran when necessary for performance reasons). PyClaw uses much of the same library of Fortran code, but that code is compiled during installation so that it can be imported dynamically within Python programs.

Algorithmic differences

All of the Clawpack solvers include the classic algorithms described in [LeVeque-FVMHP]; if you only require those, it’s easiest to use Classic or PyClaw. Most of the packages contain additional algorithms:

  • AMRClaw includes block-structured adaptive mesh refinement that allows one to use a non-uniform grid that changes in time and uses smaller grid cells in regions with fine structure or where high accuracy is required.
  • GeoClaw Includes the AMR capabilities of AMRClaw and also has a number of special routines and algorithms for handling geophysical problems, including special well-balanced, positivity-preserving shallow water solvers.
  • PyClaw includes the high-order WENO-RK algorithms of SharpClaw, described in [KetParLev13].

Parallel computing

  • AMRClaw, GeoClaw, and Classic can be run in parallel using shared memory via OpenMP.
  • PyClaw can be run in parallel on distributed-memory machines using MPI (through PETSc) and has been shown to scale to tens of thousands of cores.