[MITgcm-support] AIX and numerical diffusion

[Constantinos Evangelinos]

On Thu 26 Apr 2007 15:05, Haferman, Jeffrey (Jeff) (CIV) wrote:
> I recently copied an MITgcm model that has been running fine compiled
> under FC3 Linux with 4 dual core AMD 865 processors (i.e. 8 processing
> elements total) to an IBM p690 running AIX 5.2 (power 4 chip), also
> using 8 processors.  For the IBM I used a modified version of the
> build_options file called "sp4+mpi" supplied with MITgcm, and things
> compiled just dandy.
>
>
>
> I guess I should probably go back and run some of the test cases
> supplied with MITgcm, but, what I am seeing with my model is that I have
> pretty good agreement with the model output (T, S, u,v,w) between the
> Linux and IBM runs for the first couple of hundred time steps, though I
> see differences in some of the least significant digits.  After a few
> hundred timesteps, the model "blows up" numerically on the IBM, but
> converges fine on the Linux machine.  I am using identical input for
> both cases.

That smacks of the compiler messing things up on the IBM.

> I guess I can try to decrease the timestep on the IBM to see if it
> helps, but does has anyone else had this type of experience when moving
> a model to a new platform, and can you offer some suggestions?  I am
> using the IBM xlf and xlc compilers under their POE environment, if that
> makes any difference.

Please rerun genmake2 with the -ieee flag, rebuild and run and let us know if 
you experience the same problem. We may need to change the default 
optimization level on the IBMs.

Constantinos
-- 
Dr. Constantinos Evangelinos
Department of Earth, Atmospheric and Planetary Sciences
Massachusetts Institute of Technology