[MITgcm-devel] testreport

[Jean-Michel Campin]

Hi Martin,

Thanks for looking at this.

There can be few things which can help to get a better "score":
- advection tests & offline tests: adding a file tr_checklist
 to test something else than cg2d (like advect_cs).
 I am planning to add few tr_checklist once the AD testing is set-up.
- And when the testreport is run, can set a lower criteria to 
 get a "pass" (with -match option). The default is 13,
 but a lower criteria of 9 was used before (instead of counting the 
 number of pass) when building the testing web page.
I would propose to use 10 instead of 13 on any non linux
platform as well as with compilers which troncate differently
from g77. Note that the pgf test on faulks is not currently
giving a good score neither.

Otherwise, I checked the differences in summary.txt
between the recent one (Sept 16) and the one earlier
this month (Sept 02), directly in the directory where those
results are stored:
/net/orwell/export/export-9/mitgcm-testing/results/2007_09
And apart from few differences due to code changes, did not
see much differences for the test on edvir, xd1, rays & model.

But at the end, you are right in that few experiments don't
behave very well on different platforms, but simple things
like "_d 0" (e.g., in KPP) could make a (big) difference.

Cheers,
Jean-Michel 

On Mon, Sep 17, 2007 at 08:24:04AM +0200, Martin Losch wrote:
> Hi there,
> 
> all my tests this weekend ran with interpretable results due to Jean- 
> Michel's fix(es) to testreport. Thank you!
> 
> Just a comment: I get a lot of "FAIL"s on non-g77/non-linux machines  
> (see, e.g the machines model and rays1 are Suns, edvir1 is an IBM  
> p690, xd1 is a Cray XD1, even with Linux). I don't normally test my  
> changes on more than one machine either, but could you have a quick  
> look through the summary and tell me which of your recent changes may  
> have caused the many "FAIL"s, maybe I have an idea for a more stable  
> solution?
> 
> Martin
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel