PARDISO

[JohnNichols]

I have been using PARDISO for a while and it is really good. 

I have the solver linked up to an old Fortran program that generates a standard stiffness matrix for a beam. 

I was looking at the output for a timber beam that is modelled as 4 uniform lengths -- I get the first four modes as the same value, but if I perturb the lengths slightly I get 4 different values that are now closer to the measured values provided I play with the constants a bit. 

Interesting problem, I suppose I should print out the structures matrix and look for some form of symmetry that is providing the first answer. 

Any thoughts would be appreciated? 

John

[Ying_H_Intel]

Hi Nichols,

As i understand, there are some factors may be considering here.

1.  the nature of problem, which may be disturbed by slightly different of input. which may related your models.

2. about floating point computation and Numerical Reproducibility .  in MKL we had introduce one CNR functionality, please see MKL user guide or

the on-line article.  https://software.intel.com/en-us/articles/introduction-to-the-conditional-numerical-reproducibility-cnr

If considering the CNR, you may try the functionality and let see if any change.

Best Regards,

Ying

[JohnNichols]

Dear H Ying:

Thanks for the response.  No -- the problem is the beam is 2 metres long, the beam theory model is set up with 0.5 m long beam elements. If I run FEAST on the structures and mass matrix as an eigenvalue problem then I get eigenvalues in groups of 3 - you can see it in the picture, although here I have perturbed the lengths slightly so they are not all exactly the same -- in reality no stochastic process has perfect means and zero standard. 

It is reasonably easy with FEAST to work out which equations in the structure apply to each eigenvalue.  I am reading this correctly. 

John

[Ying_H_Intel]

Hi Jonh,

Could you provide us the input sparse matrix and the parameters you use when call PARDISO so we can debug the exact issue?

Best Regards,

Ying