[petsc-users] Inaccuracy while calculating numbers on different machines 32-bit and 64-bit.

[Matthew Knepley]

On Wed, Mar 21, 2012 at 11:05 AM, Markus Iivonen <
Markus.Iivonen at metropolia.fi> wrote:

> I've read that inaccuracy is quite common problem but on the other hand my
> values aren't that small that there should be any inaccuracy at least I
> think so...
>
> So the problem is that when I run the code on the other machine it causes
> inaccuracy which happens as early as 0.00000 digits.
>
> My program calculates two different values at the same time. First value
> is really accurate and it is calculated by boost bisection the second value
> is produced by two different functions which makes matrix and vector
> calculations using PETSc. Here are some example values:
>

Lists of numbers are meaningless. I guarantee you that these do not result
from roundoff. I will bet
that you are running some sort of solve, which is sensitive to the
tolerance.

Send a simple example to petsc-maint at mcs.anl.gov.

   Matt


> Where to compare:
>
> 0.01985308798     -0.8497140415
> 1                          0.3683436681
> 0.509926544        -0.2559051838
> 0.754963272         0.06256485518
>
> Here is what my one computer produces:
>
> 0.0198531   -0.849714
> 1                 0.368344
> 0.509927    -0.255905
> 0.754963     0.0625649
>
> And here is what happens when I'll run the program on other machine.
>
> 0.0198531      -0.849714
> 1                    0.368321
> 0.509927       -0.255922
> 0.754963        0.062544
>
> My one machine is running ubuntu 32-bit and other is running 64-bit. The
> actual values where I'm comparing to are calculated on the 64-bit system
> but different program.
>
> If I run the example value program which generated my base values on my
> 32-bit desktop it will give the same values as my C++ program and the
> example program on 64-bit system. When I run the C++ program on the 64-bit
> system the inaccuracy occurs. The values that I'm using are always doubles.
>
> Somehow the inaccuracy seems to happen mostly in the values that are
> produced by PETSc matrix and vector operations.
>
> Compilers, libraries and library confs are exactly same on the both
> machines. I did also try to set locales but it didn't have any affect, I
> just searched something and found this about PetscInt:
>
> PETSc type that represents integer - used primarily to represent size of
> arrays and indexing into arrays. Its size can be configured with the option
> --with-64-bit-indices - to be either 32bit or 64bit [default 32 bit ints]
>
> So should I maybe conf 64-bit machine with --with-64-bit-indices or should
> I try to find the problem somewhere else ?
>
> Thank You.




-- 
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/20120321/b534efa1/attachment.htm>