Weired issues when changin linear_solver_maxiter (adaptive timestepping).

[blattms]

I have seen quite unexpected behaviour in the adaptive time stepping before (when I tried potentially better parallelization that made the time stepping break down). As it came to my attention that flow_ebos uses a different default for the maximum number of linear iterations I could not help experimenting a bit with it.

Norne

Using flow_ebos with linear_solver_maxiter equal to either 75 or 150 the number of linear and nonlinear iterations, linearization, and the total runtime stays the same.

Using flow_legacy with linear_solver_maxiter equal to 150 the adaptive time stepping breaks down at step 150. This is totally unexpected! Any ideas?

Model 2

For flow_legacy the the two runs are very similar

max	Time	Linearizations	Nonlin steps	lin steps
150	8611	2332	1936	54204
75	8866	2332	1936	54204

For flow_ebos the performance drops drastically if we allow less linear steps

max	Time	Linearizations	Nonlin steps	lin steps
150	5371	2068	1724	46382
75	7254	2637	2085	51501

I assume the nonlinear steps are not counted here if the nonlinear solver does not converge. Thus the number of linearizations is actually the total number of nonlinear steps including non-convergent runs. It seems like for the flow_ebos the quality of the linear solver is of much more importance. What makes me wonder is that the behavior differs for different models.

[alfbr]

The main difference in terms of linear solver is that the ebos based one preconditions the system with only cells (not wells). I assume this has more of an impact on model 2 since some of the wells there are longer and connect larger parts of the system.

[blattms]

I am closing this and in part apologize for any inconvenience.

I realized that for flow_legacy I used a version before the default maximum number of steps was changed. That explains the why there is no change in the number of linear iterations for Model 2 above ( as nothing changed in max_iter despite what the first column says).

I was also unable to reproduce the non-convergence behavior.