Ok, thanks again,
I am a very young petsc user and even though I am familiar with MPI I'll risk myself doing something
like that. I'll stick with the easy-less-memory-efficient method for now.
cheers
aurelien

Le 17/09/16 à 20:24, Barry Smith a écrit :
    Understood. We don't have any direct way with DA for storing this 
information directly associated with you 3D DA.

     You need to figure out how to store it so that each process has access to 
the parts of the data that it needs; which may not be completely trivial. You 
could possibly use some 2d DMDA on sub communicators with suitable layouts to 
be accessible, you need to figure out the details., the dx[k] probably can be 
just stored on every process since it is 1d.

   Barry

On Sep 17, 2016, at 2:58 AM, Aurelien Ponte <aurelien.po...@ifremer.fr> wrote:

Thanks Barry for your answer !

I guess my concern is of the second type:
By grid metric terms I meant essentially grid spacings which look like:
dx(i,j), dy(i,j) and dz(k) where (i,j,k) are indices running along the 3 
dimensions of the grid.
Storing dx(i,j) into a 3D array seemed like a bit waste of memory to me but I 
must
be wrong. The elliptic problem I am solving for is close to a poisson equation 
btw.
I guess I can at least store dx and dy into a single dxy 3D array.

Thanks again,

Aurelien



Le 16/09/16 à 19:43, Barry Smith a écrit :
On Sep 16, 2016, at 9:29 AM, Aurelien PONTE <aurelien.po...@ifremer.fr> wrote:

Hi,

I've started using petsc4py in order to solve a 3D problem (inversion of 
elliptic operator).
I would like to store 2D metric terms describing the grid
      What do you mean by 2D metric terms describing the grid?

      Do you want to store something like a little dense 2d array for each grid 
point? If so create another 3D DA with a dof = the product of the dimensions of 
the little dense 2d array and then store the little dense 2d arrays at in a 
global vector obtained from that DA.

     Or is the grid uniform in one dimension and not uniform in the other two and hence you want to 
store the information about the non-uniformness in only a 2d array so as to not "waste" 
the redundant information in the third direction? Then I recommend just "waste" the 
redundant information in the third dimension; it is trivial compared to all the data you need to 
solve the problem.

     Or do you mean something else?

   Barry
I am working on but don't know
how to do that given my domain is tiled in 3D directions:

self.da = PETSc.DMDA().create([self.grid.Nx, self.grid.Ny, self.grid.Nz],
                                      stencil_width=2)

I create my 3D vectors with, for example:

self.Q = self.da.createGlobalVec()

What am I supposed to do for a 2D vector?
Is it a bad idea?

thanks

aurelien


--
Aurélien Ponte
Tel: (+33) 2 98 22 40 73
Fax: (+33) 2 98 22 44 96
UMR 6523, IFREMER
ZI de la Pointe du Diable
CS 10070
29280 Plouzané



--
Aurélien Ponte
Tel: (+33) 2 98 22 40 73
Fax: (+33) 2 98 22 44 96
UMR 6523, IFREMER
ZI de la Pointe du Diable
CS 10070
29280 Plouzané

Reply via email to