[petsc-users] About DIVERGED_ITS

[Mark F. Adams]

On Jul 10, 2012, at 10:51 AM, w_ang_temp wrote:

>  
>     In my opinion, convergence in PETSc is decided by rtol, atol and dtol. The divergent hints just show that
>  
> in the solving process it does not satisfy the rule. The "right" result may be different from the true result at
>  
> the several back decimal places(I mean that they may be the same with four decimal places but may be not
>  
> the same with more decimal places). 

No, If you have an rtol=1.e-30 and atol=1.e-300 then your solution will probably be correct to all decimal places even though PETSc will say you "diverged".  You can't be too sure about how many digits of accuracy you have with simple linear algebra arguments unless your system is extremely well conditioned.

As Jed said if you are doing 10000 iterations are happy with the solution then your tolerances are probably too tight.  Look at the residual history and see if you are stagnating.

>    
>     Is it right?
> 
> 
> >At 2012-07-08 00:28:54,"Mark F. Adams" <mark.adams at columbia.edu> wrote:
> >It sounds like your -ksp_rtol is too small.  Experiment with looser tolerances until your solution is not "correct" to see >how much accuracy you want.
> 
> >On Jul 7, 2012, at 12:15 PM, w_ang_temp wrote:
> 
>> >    Maybe it is a problem of mathematical concept. I compare the result with the true result which is
>>  
>> >computed and validated by other tools. I think it is right if I get the same result.
>> 
>> >>在 2012-07-08 00:03:21，"Matthew Knepley" <knepley at gmail.com> 写道：
>> >>On Sat, Jul 7, 2012 at 10:00 AM, w_ang_temp <w_ang_temp at 163.com> wrote:
>> >>Hello,
>>  
>> >>    I am a little puzzled that I get the right result while the converged reason says that 'Linear solve >>did not
>>  
>> >>converge due to DIVERGED_ITS iterations 10000'. This infomation means that the iterations >reach >the maximum
>>  
>> >>iterations. But the result is right now. So why says 'did not converge'? Can I think that the result is >>right and
>>  
>> >>can be used?
>> >>Obviously, your definition of "right" is not the same as the convergence tolerances you are using.
>> 
>> >>    Matt
>>  
>>  
>> >>    Thanks.
>>    
>> >>                                            Jim
>> 
>> 
>> 
>> 
>> 
>> -- 
>> >What most experimenters take for granted before they begin their experiments is infinitely more >interesting than any results to which their experiments lead.
>> >-- Norbert Wiener
>> 
>> 
> 
> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120710/5eb29314/attachment.html>