TPSA restructure (3/3): Newton defaults for linear elasticity - #7359
TPSA restructure (3/3): Newton defaults for linear elasticity#7359svenn-t wants to merge 1 commit into
Conversation
svenn-t
added
the
manual:enhancement
|
jenkins build this please |
fc2e712 to
b57acf1
Compare
b57acf1 to
ee0b4f4
Compare
|
The previous PRs have been merged. Setting this to ready-for-review. |
ee0b4f4 to
f01b307
Compare
f01b307 to
21f6512
Compare
|
jenkins build this serial please |
|
The jenkins errors are with OPMEXTRA. I'm not really sure what that is even. Something about time stepping? |
|
yes, OPMEXTRA[0] is suggested next step size i believe. |
Ok, thanks for the answer! I cannot replicate the error locally, so I just rebased. Hope that it goes through now, if not I will look into it further... |
|
jenkins build this please |
|
even more fun, only a single test failed now. this will need some sanitizers /valgrind to figure out it out, smells like some uninitialized variable or thereabout. |
I'm on the case. Will try to track it down with valgrind/debugging... |
- newton procedure now: check linearization, if error larger than tolerance, solve linear system - lowered linear solve tolerance - more clear fixed-stress conv. check - bug fix: num. dofs on owned cells only - moved calc. of info to mech. output to separate function, which is called at endTimeStep() - option to scale linear system for better conv. check with dune solvers (preconditioners are invariant)
21f6512 to
9ca0d02
Compare
I reproduced the error locally, and it seems the tighter linear solver tolerances (to compensate for only doing 1 Newton iter.) and introducing matrix/vector weights, lead to small changes in the solution which again introduces a tiny error in OPMEXTRA compared to master. If I loosen the linear solver parameters and go back to no weighting of matrix/vector, the error goes away. |
2 participants
Change Newton to default to 1 iteration, to speed up linear elasticity solves with TPSA. Now we only do one linearization, check the error, and proceed with one linear solve if linearization error is too high. Increased linear solver tolerance to compensate for only doing 1 Newton iteration.
Also includes some minor changes and fixes:
Includes both PRs #7357 and #7358, which should be merged first. Draft mode until then.