[MITgcm-devel] global_ocean.90x40x15 on sx8 fails

[Jean-Michel Campin]

Hi Martin,

I wanted to have blank-tiles tested in a more realistic set-up 
than just the adjustment-cs, but then the number of tiles is no
longer a nice multiple of small number (and in particular not even);
in addition, I though it would be nice to really test that the results
do not change, by comparing a non-mpi reference run (with no blank-tile)
with an MPI run which has 1 blank-tile.

Now, I am running a couple of testreport on more than 2 proc,
and for those tests, this experiment run on 5 proc. 
And for the more traditional testing (such as on beagle and pleiades)
which uses no more than 2 procs, using a flexible command 
(with TR_NPROC beeing replaced by the number of proc
and TR_MFILE beeing replaced with a potentially shorter version
of MACHINE_FILE), it will default to just 1 proc.
I put some explanation some time ago (last April) in the HOWTO document,
but global_ocean.90x40x15 is the 1rst exp which does not
run on 2 procs.

I cc this to devel list (might be of interest for other testreport users).

Cheers,
Jean-Michel

On Mon, Jan 09, 2012 at 10:30:49AM +0100, Martin Losch wrote:
> Hi Jean-Michel,
> 
> global_ocean.90x40x15 fails on sx8:
> (PID.TID 0000.0001) *** ERROR *** S/R EEBOOT_MINIMAL: No. of processes not equal to nPx*nPy     2     1
> 
> I guess this is intentional, but still I'd like to understand what's going on. In 
> input/SIZE.h_mpi
> [...]
>       PARAMETER (
>      &           sNx =  10,
>      &           sNy =  10,
>      &           OLx =   3,
>      &           OLy =   3,
>      &           nSx =   7,
>      &           nSy =   1,
>      &           nPx =   1,
>      &           nPy =   5,
>      &           Nx  = sNx*nSx*nPx,
>      &           Ny  = sNy*nSy*nPy,
>      &           Nr  =  15)
> [...]
> 
> Nx=70, Ny=50 > 3500 surface grid point, but 90x40=3600 surface grid nodes. Is this the missing tile and you can arrange the tiles in any way you like? Pretty cool.
> 
> How do you explain this?
> cat build/SIZE.h
> [...]
>       PARAMETER (
>      &           sNx =  10,
>      &           sNy =  10,
>      &           OLx =   3,
>      &           OLy =   3,
>      &           nSx =   7,
>      &           nSy =   5,
>      &           nPx =   1,
>      &           nPy =   1,
>      &           Nx  = sNx*nSx*nPx,
>      &           Ny  = sNy*nSy*nPy,
>      &           Nr  =  15)
> [...]
> 
> Do I have to (Can I?) do anything to make the experiment run on SX8 (other than asking for 5 processors, which is difficult)?
> 
> Martin
>