It may have been captured elsewhere already, but using the test example provided “beamb.inp” runnning with Spooles solver and 4 processors gives the following buckle factors for the 1st mode when repeating the step 10 times.
4.82E+01
1.75E+00
4.82E+01
1.88E+00
4.82E+01
4.82E+01
4.82E+01
4.82E+01
4.82E+01
4.82E+01
repeating the analysis does not give consistency either. Issue seems to be unique to spooles and OMP_NUM_THREADS >2
Today, I reported an issue, part of which is also that Spooles provides inconsistent results in each run when more than 2 cpus are used: First buckling mode skipped when applied load is much bigger than buckling load · Issue #74 · Dhondtguido/CalculiX · GitHub
2 Likes
Multithreaded Spooles (at least in CCX) is long known to have these bugs. I think nobody should ever use it.
Correct solvers:
- Single threaded Spooles
- MKL Pardiso
- Pastix with mixed precision turned off by an environment variable
Wrong solvers:
- Multithreaded Spooles
- Pastix with mixed precision turned on (default)
2 Likes