Distortion with S6 elements when using ccx > 2.17 and solvers other than Pardiso

[FEA-eng]

I've seen similar issues in the case of bars meshed with tetrahedral elements but only in buckling/frequency analyses. This model is a thin-walled closed cylinder meshed with S6 elements. It's loaded with gravity, bending force at the free end and external pressure on the whole cylindrical surface. The results are incorrect when using CalculiX newer than 2.17 unless Pardiso is selected:

It also works with first-order elements.

Here's the input file: https://www.dropbox.com/scl/fi/7q9eu4lqqrtrcrn413vki/FEMMeshGmsh.inp?rlkey=4eo15kp9gd71q7n51029506j7&st=o7o9cagn&dl=0

[victorkemp]

This is because of mixed precision with Pastix. Use the environment variable PASTIX_MIXED_PRECISION = 0 to turn it off.

The output shows Pastix finishing its iterations with a large error:

• iteration 48 :
  total iteration time 0.0593
  error 0.074056
• iteration 49 :
  total iteration time 0.065
  error 0.074051
• iteration 50 :
  total iteration time 0.0608
  error 0.074004

With mixed precision off, error is near zero:

• iteration 1 :
  total iteration time 0.0764
  error 3.0545e-12
  Time to solve: 0.0685
• iteration 2 :
  total iteration time 0.0776
  error 1.7296e-12
  Time to solve: 0.0682
• iteration 3 :
  total iteration time 0.0772
  error 1.7231e-13

[FEA-eng]

This is because of mixed precision with Pastix. Use the environment variable PASTIX_MIXED_PRECISION = 0 to turn it off.

Thanks. Do you think that it should be set this way by default or are there some issues with it?

[victorkemp]

I do think it should be the default because mixed precision causes so many errors. The only problem I know of is that it's probably slower.