[MITgcm-devel] Re: [MITgcm-support] Question: boundary exchange, hrcube condfiguration

[Martin Losch]

Hi Jean-Michel (and Chris!),

I still don't understand, your argument would apply to non-cube  
topologies too, but it does not.

In exch2_3d_r8 we have:
        CALL EXCH2_R81_CUBE( phi, 'T ',
      I            OLw, OLe, OLs, OLn, myNz,
      I            exchWidthX, exchWidthY,
      I            FORWARD_SIMULATION, EXCH_UPDATE_CORNERS, myThid )
then there a block within
#ifdef W2_FILL_NULL_REGIONS
#endif
which according to some comment is only used for debugging. defining  
W2_FILL_NULL_REGIONS fills the face corners with zeros (or whatever  
the fill value is) and this is not supposed to change the results, so  
why do we need a second call (after which filling the corners does/ 
should not change the results)?

Martin

On 27 Jun 2007, at 15:24, Jean-Michel Campin wrote:

> Hi,
>
> I think (but I might be wrong) that,
> in the simplest case (a tile in the middle of a face),
> we need a 1rst call to fill the halo region next to the tile edges
> (i.e., [1-Olx:0,1:sNy]),
> and a second call to fill the "corner" parts of the halo regions
> (i.e, [1-Olx:0,1-Oly:0]).
>
> This could be done in just 1 call, but would require to connect to
> 8 tiles instead of only 4 in the present case.
>
> I would be interesting to see how to apply the same solution
> as with exch1 (or why it's not possible).
>
> Now, I am trying to figure out why we need 3 exit2_rl2_cube calls ...
>
> Jean-Michel
>
> On Wed, Jun 27, 2007 at 01:44:54PM +0200, Martin Losch wrote:
>> Hi there,
>>
>> could someone please explain this to Michael (and me)? Why do we need
>> two calls to exch2_rl1_cube?
>>
>> Martin
>>
>> On 22 Jun 2007, at 14:14, Michael Schroeter wrote:
>>
>>> Hi,
>>>
>>> as Martin already reported in an earlier posting
>>> (http://forge.csail.mit.edu/pipermail/mitgcm-support/2007-March/
>>> 004719.html)
>>> we have some unexpected scaling behaviour on a IBM p690
>>> (cubed sphere configuration, 16 passive tracers).
>>>
>>> We found out that the boundary exchange of the ptracers data seems
>>> to be
>>> responsible for the slow down of the code. Now we have some ideas  
>>> for
>>> code optimazation.
>>>
>>> However, there is one thing in the current code I do not unterstand.
>>> In "exch2_3d_rl.F" the subroutine "EXCH2_RL1_CUBE" is called two  
>>> times
>>> with exactly the same values for the argument list parameters
>>> (even if the cpp option "W2_FILL_NULL_REGIONS" is undefined).
>>> Can anybody explain that to me?
>>>
>>> Thanks in advance.
>>>
>>> Regards
>>> Michael
>>>
>>> -- 
>>> Dr. Michael Schroeter                          Phone: +49(471)
>>> 4831-2084
>>> Email: Michael.Schroeter at awi.de                Fax:   +49(471)
>>> 4831-1149
>>> Alfred-Wegener-Institute for Polar- and Marine Research
>>> Am Handelshafen 12, D-27570 Bremerhaven
>>> www.awi.de
>>> _______________________________________________
>>> MITgcm-support mailing list
>>> MITgcm-support at mitgcm.org
>>> http://mitgcm.org/mailman/listinfo/mitgcm-support
>>
>> _______________________________________________
>> MITgcm-devel mailing list
>> MITgcm-devel at mitgcm.org
>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel