Code Organization

The family of solvers referred to as SUNDIALS consists of the solvers CVODE and ARKode (for ODE systems), KINSOL (for nonlinear algebraic systems), and IDA (for differential-algebraic systems). In addition, SUNDIALS also includes variants of CVODE and IDA with sensitivity analysis capabilities (using either forward or adjoint methods), called CVODES and IDAS, respectively.

The various solvers of this family share many subordinate modules. For this reason, it is organized as a family, with a directory structure that exploits that sharing (see the following Figures SUNDIALS organization and SUNDIALS tree). The following is a list of the solver packages presently available, and the basic functionality of each:

Note for modules that provide interfaces to third-party libraries (i.e., LAPACK, KLU, SuperLU_MT, SuperLU_DIST, hypre, PETSc, Trilinos, and RAJA users will need to download and compile those packages independently.


SUNDIALS organization: High-level diagram of the SUNDIALS structure


SUNDIALS tree: Directory structure of the source tree.

ARKode organization

The ARKode package is written in the ANSI C language. The following summarizes the basic structure of the package, although knowledge of this structure is not necessary for its use.

The overall organization of the ARKode package is shown in Figure ARKode organization. The central integration modules, implemented in the files arkode.h, arkode_impl.h, arkode_butcher.h, arkode.c, arkode_arkstep.c , arkode_erkstep.c, arkode_mristep.h, and arkode_butcher.c, deal with the evaluation of integration stages, the nonlinear solvers, estimation of the local truncation error, selection of step size, and interpolation to user output points, among other issues. ARKode currently supports modified Newton, inexact Newton, and accelerated fixed-point solvers for these nonlinearly implicit problems. However, when using the Newton-based iterations, or when using a non-identity mass matrix \(M\ne I\), ARKode has flexibility in the choice of method used to solve the linear sub-systems that arise. Therefore, for any user problem invoking the Newton solvers, or any user problem with \(M\ne I\), one (or more) of the linear system solver modules should be specified by the user, which is then invoked as needed during the integration process.


ARKode organization: Overall structure of the ARKode package. Modules specific to ARKode are the timesteppers (ARKODE), linear solver interfaces (ARKLS), nonlinear solver interfaces (ARKNLS), and preconditioners (ARKBANDPRE and ARKBBDPRE); all other items correspond to generic SUNDIALS vector, matrix, and solver modules.

For solving these linear systems, ARKode’s linear solver interface supports both direct and iterative linear solvers built using the generic SUNLINSOL API (see Description of the SUNLinearSolver module). These solvers may utilize a SUNMATRIX object for storing Jacobian information, or they may be matrix-free. Since ARKode can operate on any valid SUNLINSOL implementation, the set of linear solver modules available to ARKode will expand as new SUNLINSOL modules are developed.

For users employing dense or banded Jacobians, ARKode includes algorithms for their approximation through difference quotients, although the user also has the option of supplying a routine to compute the Jacobian (or an approximation to it) directly. This user-supplied routine is required when using sparse or user-supplied Jacobian matrices.

For users employing iterative linear solvers, ARKode includes an algorithm for the approximation by difference quotients of the product \(Av\). Again, the user has the option of providing routines for this operation, in two phases: setup (preprocessing of Jacobian data) and multiplication.

When solve problems with non-identity mass matrices, corresponding user-supplied routines for computing either the mass matrix \(M\) or the product \(Mv\) are required. Additionally, the type of linear solver module (iterative, dense-direct, band-direct, sparse-direct) used for both the IVP system and mass matrix must match.

For preconditioned iterative methods for either the system or mass matrix solves, the preconditioning must be supplied by the user, again in two phases: setup and solve. While there is no default choice of preconditioner analogous to the difference-quotient approximation in the direct case, the references [BH1989] and [B1992], together with the example and demonstration programs included with ARKode and CVODE, offer considerable assistance in building simple preconditioners.

ARKode’s linear solver interface consists of four primary phases, devoted to

  1. memory allocation and initialization,
  2. setup of the matrix/preconditioner data involved,
  3. solution of the system, and
  4. freeing of memory.

The setup and solution phases are separate because the evaluation of Jacobians and preconditioners is done only periodically during the integration process, and only as required to achieve convergence.

ARKode also provides two rudimentary preconditioner modules, for use with any of the Krylov iterative linear solvers. The first, ARKBANDPRE is intended to be used with the serial or threaded vector data structures (NVECTOR_SERIAL, NVECTOR_OPENMP and NVECTOR_PTHREADS), and provides a banded difference-quotient approximation to the Jacobian as the preconditioner, with corresponding setup and solve routines. The second preconditioner module, ARKBBDPRE, is intended to work with the parallel vector data structure, NVECTOR_PARALLEL, and generates a preconditioner that is a block-diagonal matrix with each block being a band matrix owned by a single processor.

All state information used by ARKode to solve a given problem is saved in a single opaque memory structure, and a pointer to that structure is returned to the user. For C and C++ applications there is no global data in the ARKode package, and so in this respect it is reentrant. State information specific to the linear solver interface is saved in a separate data structure, a pointer to which resides in the ARKode memory structure. State information specific to the linear solver implementation (and matrix implementation, if applicable) are stored in their own data structures, that are returned to the user upon construction, and subsequently provided to ARKode for use. We note that the ARKode Fortran interface, however, currently uses global variables, so at most one of each of these objects may be created per memory space (i.e. one per MPI task in distributed memory computations).