[MITgcm-devel] new r4 SingleCpuIO efficiency

[Jean-Michel Campin]

Hi Dimitris,

The 2nd isssue is easier: I think (would need to check) that
Olx needs to be smaller than sNx and same thing for Oly,sNy.
With exch-1, we ended having 2 special case for 2-D set-up
(x-z or y-z) for this reason (if I remember well).
Now, for a simple set-up (no seaice, Hydrostatic ...),
compiled with zero optimisation, the GLOBAL_SUM_SINGLECPU
will help to get same results independent of the tiling.

The 1rst issue: the check-depth is on the "safe" side:
if there is zero depth at the edge of a tile touching a blank-tile,
then it's relatively clear that it is OK.
I don't know if this is safe to allow a non-zero depth 
next to a blank-tile (because it might depend of which piece of 
the algorithm we consider). But it could well be that I am
wrong, that the default initialisation provides a "safe" set-up.
But regarding your check of hfacW,S, ... , does this includes
the overlap ?

Cheers,
Jean-Michel

On Mon, May 25, 2009 at 05:48:49PM -0700, Dimitris Menemenlis wrote:
> Jean-Michel, I was trying to recover an old test experiment on the cs32 
> grid
> with blank tiles and I get some errors of type:
>
> (PID.TID 0000.0001) *** ERROR *** EXCH2_CHECK_DEPTHS: tile #    13  
> (bi,bj=  13,   1 ):
> (PID.TID 0000.0001) *** ERROR *** W.Edge has    1 unconnected points  
> with non-zero depth.
>
> etc., and the program aborts.
>
> If I comment exch2_check_depths, the code will run OK and it will  
> produce hFacW.data,
> hFacS.data, hFacC.data, and Depth.data output files that are exactly  
> identical to those
> obtained without blank tiles, so I am pretty sure that blank tiles are  
> chosen and specified correctly.
>
> Is there a requirement that halos of blank tiles also be land points?
> Or is there a restriction pertaining to OLx/Oly compared to sNx/sNy?
>
> ===
>
> A separate, possibly related issue, integrating a cs32 experiment with  
> different tile configurations, I get substantial differences in ETAN  
> after 72 time steps, especially when OLy<=sNy.
> For example,
> 24-tile vs 12-tile setup: difference is < real*4 precision of output  
> files
> 176-tile vs 192-tile: up to 4e-6 m difference
> 176-tile vs 12-tile: up to 4e-2 cm difference
>
> Is there some restriction pertaining to OLx/Oly compared to sNx/sNy in  
> order for the
> solver to be accurate?
>
> ===
>
> If helpful, the test set-up that I am experimenting with is documented  
> in
> the first two experiments described in
> MITgcm_contrib/high_res_cube/README.cs32
>
> D.
>
> On May 22, 2009, at 11:09 AM, Jean-Michel Campin wrote:
>
>> If you are not using blank-tiles, then you just need to remove
>> our local copy of w2_e2setup.F & W2_EXCH2_TOPOLOGY.h
>> With blank-tiles, you remove also thoses 2 files but you need
>> a "data.exch2" parameter file (with the list of blank-tiles),
>> and could find one example (with blank-tiles) here:
>> MITgcm/verification/adjustment.cs-32x32x1/input/data.exch2.mpi
>> (the .mpi suffix is just there for MPI testing purpose).
>> If this does not work, you should complain.
>
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel