[MITgcm-devel] verification/lab_sea/tr_run.salt_plume

[Jean-Michel Campin]

Hi Dimitris,

Let's reserve the word "stable" for numerical stability issue,
and talk about "sensitivity" of this set-up when results
change from 1 platform/compiler to an other.

In your previous e-mail: 
Advection scheme 7 for T & S => not using AB for T & S 
  => you need StaggerTimeStep to get a "stable" time-stepping
scheme regarding internal-wave processes. 
There are obviously some special cases where it's not needed (e.g., 
if there is no stratification), and other cases where it's not clear
(e.g., with nonHydrostatic=T). But in general, it's better to 
use a stable time-stepping.

Sensitivity of the set of lab_sea test:
1) on faulks, we only get 5 or 6 digits for cg2d with ifort or pgf.
for all 4 lab_sea tests. In this situation, it's difficult to say that 
what you get with StaggerTimeStep=T in lab_sea.salt_plume 
is very specific to StaggerTimeStep.

2) I agree, it would be usefull to know why (and which part of
the code) is responsible for this high sensitivity of all lab_sea
test. I remember trying to investigate this question but did not
get very far.

3) regarding this example below, cg2d is still better with 
StaggerTimeStep=T (13 compared to 12).

That's all for today.
Cheers,
Jean-Michel


On Wed, Apr 23, 2008 at 06:04:14PM -0700, Dimitris Menemenlis wrote:
> here is results for sensitivity test with useSALT_PLUME=.FALSE.
> 
> >useSALT_PLUME=.FALSE., AdvScheme=7, and StaggerTimeStep=.FALSE.
> >Y Y Y Y>12<16 16 16 16 16 16 16 16 16 16 13 16 16 16 13 16  .  .  .  .  .  
> >.  .  . FAIL  lab_sea.salt_plume
> >
> >useSALT_PLUME=.FALSE., AdvScheme=7, and StaggerTimeStep=.TRUE.
> >Y Y Y Y>13<16 16 16 10 16 16 10  8 16 16 13  6 13 16 13  6  .  .  .  .  .  
> >.  .  . pass  lab_sea.salt_plume
> 
> again, for AdvScheme=7, StaggerTimeStep=.FALSE. appears to be more stable 
> than StaggerTimeStep=.TRUE.
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel