After 2ea5eb1 which added a dummy element for the matrixbased Stokes implementation, I noticed a different hp-refinement behavior in the 3D ypipe test. Interestingly, the 2D Kovasznay test remains unaffected.
Try to...
...recreate error with pure h-refinement to rule out errors in the SmoothnessEstimator
...then compare the number of cells refined to rule out errors in the KellyErrorEstimator
...check matrix/solution norms
Also investigate any effects on the solver performance. The AMG version of both matrixbased and matrixfree implementations should produce the same result.
Test: $ ctest -V -R ypipe_dealiitrilinos.prm
New behavior:
Cycle 0:
Number of global levels: 1
Number of active cells: 72
Max active FE index in use : 1
Frequencies of poly. degrees: 2:72
Number of degrees of freedom: 2420
Solved in 32 iterations.
Cycle 1:
Number of global levels: 2
Number of active cells: 156
Max active FE index in use : 2
Frequencies of poly. degrees: 2:144 3:12
Number of degrees of freedom: 7467
Solved in 41 iterations.
Cycle 2:
Number of global levels: 3
Number of active cells: 317
Max active FE index in use : 2
Frequencies of poly. degrees: 2:226 3:91
Number of degrees of freedom: 20559
Solved in 50 iterations.
Old behavior:
Cycle 0:
Number of global levels: 1
Number of active cells: 72
Max active FE index in use : 0
Frequencies of poly. degrees: 2:72
Number of degrees of freedom: 2420
Solved in 32 iterations.
Cycle 1:
Number of global levels: 2
Number of active cells: 149
Max active FE index in use : 1
Frequencies of poly. degrees: 2:139 3:10
Number of degrees of freedom: 7129
Solved in 40 iterations.
Cycle 2:
Number of global levels: 3
Number of active cells: 303
Max active FE index in use : 1
Frequencies of poly. degrees: 2:222 3:81
Number of degrees of freedom: 19246
Solved in 53 iterations.
After 2ea5eb1 which added a dummy element for the matrixbased Stokes implementation, I noticed a different hp-refinement behavior in the 3D ypipe test. Interestingly, the 2D Kovasznay test remains unaffected.
Try to...
Also investigate any effects on the solver performance. The AMG version of both matrixbased and matrixfree implementations should produce the same result.
Test:
$ ctest -V -R ypipe_dealiitrilinos.prm
New behavior:
Old behavior: