[MITgcm-devel] [MITgcm-cvs] MITgcm/tools/build_options CVS Commit

[David Ferreira]

On 4/26/10 11:41 AM, Constantinos Evangelinos wrote:
> On Saturday 24 April 2010 07:42:55 pm Jean-Michel Campin wrote:
>    
>> Hi Dimitris,
>>
>> I think that if your modified optfile degrade the results of
>> the 2+2 test, it's better to go back and leave the new option
>> commented: the reason is that it's quiet nice to be able
>> to restart exactly, and if it's no longer the case, it could
>> disturb other users. I personally don't run the 2+2 test very
>> often (apart from the standard verification tests), and it might
>> take them some time to realize that they lost this capability.
>>      
> So essentially we have an option (-mp) that works but is not going to be there
> in the Intel 12.0 compilers while its suggested equivalent option (-fp-model
> precise) is producing underflows in the EVP solver slowing it down immensely
> and at the same time a modification to -fp-model precise to avoid the
> underflows (-ftz) breaks the restart... What a mess.
>
> Constantinos
>    
I've run a test report on Columbia and yes removing -mp breaks lots of 
restarts (nothing
surprising that's why we put it there in the first place and the 
compiler hasn't change).
Additionally, this completely changes the test report (and I guess any 
runs done on Columbia).

So I vote to put the -mp back for now.
And maybe it's not really a problem: on pleiades with ifort 11, -mp is 
not needed.
david