[petsc-users] Running the same problem multiple time, but the matrix is not the same

[Matthew Knepley]

On Thu, Dec 8, 2016 at 5:10 PM, Khai Pham <khaipham at utexas.edu> wrote:

> Hi Hong,
>
> I more concern about the difference between A1 and A2 is in the order of
> O(1.e-8) as I run the same code twice. Should it be in the order of machine
> epsilon?
>

If you have 17 orders of magnitude difference between elements, then its
easy to cancellation when doing subtraction and have
differences on that order due to permuted operations.

  Matt


> Khai
>
> On Thu, Dec 8, 2016 at 4:54 PM, Hong <hzhang at mcs.anl.gov> wrote:
>
>> Khai :
>>>
>>> Thanks for your response. The output is not the solution. They are the
>>> component of the matrix in matlab format. I would expect the difference in
>>> the order of machine precesion. The difference in solutions in two runs is
>>> in the order of 1e-5.
>>>
>>
>> If your matrices A1 and A2 have difference of O(1.e-8), then the computed
>> solution may differ by
>> Condition_number(A) * machine_epsion. Do you know cond(A)?
>>
>> Please alway send your request to petsc-maint.
>>
>> Hong
>>
>>>
>>>
>>> On Dec 8, 2016 4:01 PM, "Hong" <hzhang at mcs.anl.gov> wrote:
>>>
>>> Khai :
>>> Your solution components have values ranging from 1.e+9 to 1.e-8, and
>>> the values only differ in the order of 1.e-8, which are within
>>> computational error tolerance.
>>> I would consider all solutions same within the approximation tolerance.
>>>
>>> Hong
>>>
>>> Hello,
>>>>
>>>> I have problem with matrix assembly for linear solver KSP. I run the
>>>> the same problem with 4 processors multiple time. Using flag
>>>> -mat_view ::ascii_matlab to view the matrix. The matrix outputs are not the
>>>> same during each run. Please see the attached file for the comparison
>>>> between two runs. I checked all the input for each processor as it calls to
>>>> MatSetValues ( indices of the matrix and values ) and it's consistent all
>>>> the time. I also checked the allocation information (with flag -info) and
>>>> it looks fine. Could you give me an advice how to deal with this issue?
>>>> Thanks !
>>>>
>>>> Best,
>>>> Khai
>>>>
>>>
>>>
>>>
>>
>


-- 
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/20161208/39ff4483/attachment.html>