Diego,
your problem might be numerically unstable, that's why results might
differ between one run and an other.
floating point numbers have their own restrictions (rounding errors,
absorption, ...)
are you running single or double precision ?
if you are running single precision, you might give double precision a try.
(if your code is written in fortran, you can use the -r8 flag to treat
real (single precision) as double)
let me give you a (theoretical) example :
1 / (1.e+100 + 1 - 1.e+100) = ?
if you do this by hand, the answer is 1
now if you ask a computer using floating point numbers to do that, he
might do
1.e+100 +1 ~= 1.e+100
1.e+100 - 1.e+100 = 0
1 / 0 = Division by zero
an other classic example is
sum = 0
do i=1,n
sum = sum + 1/i
done
that might look trivial, but it is very hard to get accurate results
with a computer :
a naive approach will give you inaccurate results
bottom line, you notice differences, and that is normal.
the question is how do you compare your results and how much do they
differ ?
if you do a binary comparison of the results, it is very likely results
will differ.
if you compare a and b, and abs(a-b) / abs(a) is very low (depending on
you using single vs double precision),
then this is likely the normal behaviour.
now if this number is high, that could be a bug in your code (never say
never ...) or your algorithm might be numerically unstable (at least for
your test case)
Cheers,
Gilles
On 10/29/2015 7:58 AM, Diego Avesani wrote:
dear Damin,
I wrote the solver by myself. I have not understood your answer.
Diego
On 28 October 2015 at 23:09, Damien <dam...@khubla.com
<mailto:dam...@khubla.com>> wrote:
Diego,
There aren't many linear solvers that are bit-consistent, where
the answer is the same no matter how many cores or processes you
use. Intel's version of Pardiso is bit-consistent and I think
MUMPS 5.0 might be, but that's all. You should assume your answer
will not be exactly the same as you change the number of cores or
processes, although you should reach the same overall error
tolerance in approximately the same number of iterations.
Damien
On 2015-10-28 3:51 PM, Diego Avesani wrote:
dear Andreas, dear all,
The code is quite long. It is a conjugate gradient algorithm to
solve a complex system.
I have noticed that when a do cycle is small, let's say
do i=1,3
enddo
the results are identical. If the cycle is big, let's say do
i=1,20, the results are different and the difference increase
with the number of iterations.
What do you think?
Diego
On 28 October 2015 at 22:32, Andreas Schäfer <gent...@gmx.de
<mailto:gent...@gmx.de>> wrote:
On 22:03 Wed 28 Oct , Diego Avesani wrote:
> When I use a single CPU a get a results, when I use 4 CPU I
get another
> one. I do not think that very is a bug.
Sounds like a bug to me, most likely in your code.
> Do you think that these small differences are normal?
It depends on what small means. Floating point operations in a
computer are generally not commutative, so parallelization
may in deed
lead to different results.
> Is there any way to get the same results? is some align
problem?
Impossible to say without knowing your code.
Cheers
-Andreas
--
==========================================================
Andreas Schäfer
HPC and Grid Computing
Department of Computer Science 3
Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany
+49 9131 85-27910 <tel:%2B49%209131%2085-27910>
PGP/GPG key via keyserver
http://www.libgeodecomp.org
==========================================================
(\___/)
(+'.'+)
(")_(")
This is Bunny. Copy and paste Bunny into your
signature to help him gain world domination!
_______________________________________________
users mailing list
us...@open-mpi.org <mailto:us...@open-mpi.org>
Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
Link to this post:
http://www.open-mpi.org/community/lists/users/2015/10/27933.php
_______________________________________________ users mailing
list us...@open-mpi.org <mailto:us...@open-mpi.org> Subscription:
http://www.open-mpi.org/mailman/listinfo.cgi/users
Link to this
post:http://www.open-mpi.org/community/lists/users/2015/10/27934.php
_______________________________________________
users mailing list
us...@open-mpi.org <mailto:us...@open-mpi.org>
Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
Link to this post:
http://www.open-mpi.org/community/lists/users/2015/10/27935.php
_______________________________________________
users mailing list
us...@open-mpi.org
Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
Link to this post:
http://www.open-mpi.org/community/lists/users/2015/10/27936.php
