[petsc-users] DIVERGED_DTOL

[Barry Smith]

On Jan 16, 2013, at 7:55 AM, w_ang_temp  <w_ang_temp at 163.com> wrote:

> At 2013-01-16 03:02:32,"Barry Smith" <bsmith at mcs.anl.gov> wrote:
> >
> >> 1.701448294063e+04  >   1.e4*1.145582415879e+00  hence it declares divergence
>   Hello, Barry
>   I made some tests and it is true.  Thanks.
>  
>   But in the mannual, both version 3.2 and 3.3, the rule is:||rk||>dtol*||b||. It is ||b||, not r_0.
> My misunderstanding? Or the error in the manual?

   Error in the manual. Depending on the solver it is either the preconditioned norm or the unpreconditioned norm

>   Besides, in the mannual, both version 3.2 and 3.3, the default dtol=1.0E+5. But from the results
> in the example, it is 1.0E+4.

    Another error in the manual

> I donot know the reason. My petsc version is 3.2-p7.

   None of this should really matter. The solver has gone bad at this point regardless of specifics of how you measure it.

   Barry

>   Thanks.                            Jim
>  
> 
> >Note that at iteration 171 the preconditioned residual is 9.348832909193e-13 < 1.e-12 * 1.145582415879e+00  very good convergence.
> >
> >You seem to have set an unreasonably tight convergence criteria. In double precision you can never realistically expect to use a rtol smaller than e-12. In fact normally it is not reasonable to use more than like 1.e-8. Those extra digits don't mean anything.
> >
> >   Barry
> >
> >
> 
>