Since several linear solvers are available, a choice has to be made regarding whether to use one of the two in-house iterative solvers, the in-house direct solver or an external module. This choice can be based on a-priori available information. For example the performance evaluation presented in Section 5.5 resulted in some conclusions about which solvers are best-suited for different kinds of simulations.
However, although a solver is selected due to general a-priori considerations regarding the kind and dimension of the simulation, the solver still might fail whereas an alternative type may succeed in solving the equation system. Although care has to be taken while evaluating the results of such problematic simulations, a successful result may trigger particular improvements in the complete simulation setup. Keeping this goal carefully in mind, the user will normally select an alternative solver if one fails.
The manual selection of alternative solvers has one major drawback: if the convergence problem occurs after a long simulation time, the complete simulation has to be rerun with the alternative solver although this would not be necessary. An automatic solver hierarchy as implemented in the iteration control module of MINIMOS-NT resolves this problem. The simulation is started according to the user settings in the input-deck. Note that the user setting can also enable an automatic solver selection which is based on conclusions drawn from a performance evaluation.
If a problem occurs, the solver parameters are reset in order to adjust them for the new situation. If this does not help, the alternative iterative in-house solver is selected. This means that GMRES(M) replaces BICGSTAB and vice versa. If the problem persists, the direct in-house solver is selected instead of the iterative one. If the direct solver is not able to solve the system, the simulation will be terminated. Otherwise, the user settings are reestablished in order to use the original solver for the next step. In this hierarchy, the external modules can be integrated. Since the in-house direct solver will quickly exceed its limitations for large mixed-mode and three-dimensional simulations, the PARDISO direct solver would be a very good alternative.
An alternative possibility for selecting another solver is to use an interactive mode of the simulator. As shown for MINIMOS-NT (see Appendix C.2), the user can interrupt the simulation process and directly modify the settings in the input-deck database. These modifications include the selection of solver modules and systems.
Note that there exists a second solver hierarchy directly in the solver module. If no selection of a particular solver type is made (default), first the iterative solver is tried, followed by the direct one in case problems occur.