Opened 7 years ago
Last modified 3 years ago
#4510 new defect
Importance of residual scaling for the proper operation of the default nonlinear solver
| Reported by: | Francesco Casella | Owned by: | somebody |
|---|---|---|---|
| Priority: | critical | Milestone: | |
| Component: | Run-time | Version: | |
| Keywords: | Cc: |
Description
This test is similar to the one shown in #4395, but it uses the standard nonlinear solver, see the attached file.
Test2 is nicely scaled by itself. Test3 is badly scaled, but has all the nominal attributes needed to scale it up correctly.
Test2 runs in 0.7 seconds on my PC. Test3 takes 51 seconds. My conjecture is that if residuals are scaled as suggested in ticket:4395#comment:3 (a trivial operation) there is no reason performance on Test3 should not be about the same as on Test2.
Attachments (2)
Change History (9)
by , 7 years ago
| Attachment: | test-nonlin.mos added |
|---|
comment:1 by , 7 years ago
comment:2 by , 7 years ago
You may want to use the attached test-nonlin-v3.mos for a more comprehensive stressing of the solver.
by , 7 years ago
| Attachment: | test-nonlin-v3.mos added |
|---|
comment:4 by , 5 years ago
| Milestone: | 1.14.0 → 1.16.0 |
|---|
Releasing 1.14.0 which is stable and has many improvements w.r.t. 1.13.2. This issue is rescheduled to 1.16.0
comment:6 by , 4 years ago
| Milestone: | 1.17.0 → 1.18.0 |
|---|
Retargeted to 1.18.0 because of 1.17.0 timed release.
I have been able to turn on LOG_NLS_V. The well scaled problem converges in 5 iterations, while the badly scaled one takes over 14900 iterations (!)