[MITgcm-devel] cs510 on IBM p690

[Martin Losch]

Hi all,

I repeated my run on our XD1 (MITgcm/tools/build_options/ 
linux_amd64_pgf77+mpi_xd1) with 54 CPUs and I get exactly the same  
result, that is, stripes in x/i-direction in the U/V-fields, pointing  
to a problem in exch*_uv_* files in j/y-direction. Otherwise the  
output files between the two machines are exactly the same to working  
precisition.

So, it's not the machine, but the somehow the configuration.

I'll try Chris' suggestions.

M.

On 14 Feb 2007, at 12:08, Martin Losch wrote:

> Hi Dimitris,
>
> "Soccer ball" was my first association, too.
>
> I use Olx=Oly=4. (see SIZE.h)
>
> I don't use 27 tiles, but 27 CPUs. I use 216 tiles (it's your  
> s216_85x85 directory, but I have modified SIZE.h_54 to run on  
> nPx=27 instead of 54 CPUs, increasing nSx to 8, 8x27=216), so each  
> block is 6 tiles. But clever as you are, you immediatley spotted  
> that it is only the exchange in the j-direction, that is the  
> problem (it took me a whole day to figure that out). And to give  
> you even more information, this problem does NOT show up in Theta/ 
> Salt, but only in U and V (and consequently in Eta), in the first  
> time step. And it's only the j/y-direction that's strange. So I  
> suspect something wrong in the exch_uv_* and/or exch2_uv_*
>
> So, what can go wrong in the exchanges j-direction, that is  
> specific to IBM p690 (and xlf-compilers), and shows only up in this  
> particular configuration so far (the cs32 does not have these  
> stripes, I tried 4 tiles per face = 24 tiles)?
>
> Martin
>
> On 13 Feb 2007, at 23:57, Dimitris Menemenlis wrote:
>
>> Martin, that's a very nice design for next world cup soccer ball.
>>
>> One thing I find strange is that your figure has 36 rather than 27  
>> tiles?  Are you sure that you used the attached SIZE.h file.  And  
>> if not did you have
>>
>>      &           OLx =   4,
>>      &           OLy =   4,
>>
>> D.
>> _______________________________________________
>> 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