Re: Different results on two clusters

Hakan Lane <H.L@xxxxxxxxxxxxx> wrote:
: I am experiencing a rather unexpected phenomenon. The same code acting on
: the same indata produces different results on two separate clusters. On
: one, it is compiled with IFC 7.0 and on the other with IFC 9.0.

: Do you know why this would happen? It would not have been a big problem
: if there were only small numerical differences, but this is the case of
: the solution process eventually diverging on one of the systems. It is
: very important for me to get around this problem.

: I dont know if it is the machine precision, the compiler or the
: libraries. If you have an idea of what can be done, please help me.
: Thanks in advance.

: Regards,

: Hakan Lane

You have not defined the nature of your problem, but I would bet that it is a
chaotic dynamical system, for which truncation errors grow exponentially.
So, on two different systems, the solutions start out the same, but
eventually they totally diverge.

In the context of weather forcasting, this is called "the butterfly effect".
(If you have never heard of that metaphore, forget it.)
If the dynamical system is intrinsically unstable, as is the case for chaos,
then the behavior that you see is perfectly normal.

Leslie Ballentine
Simon Fraser University

.