Re: [Libmesh-devel] [Libmesh-users] block matrix implementation

[Lorenzo A. B. <lor...@gm...>]

I tried to modify dof_map.h as follows 
  unsigned int block_size() const
  {
#ifdef LIBMESH_ENABLE_BLOCKED_STORAGE
    return (this->has_blocked_representation() ? this->n_variables()*4 : 1);
#else
    return 1;
#endif
  }

I don't know if it is the "correct" way to do it but the memory footprint has not changed.
I'm also interested in knowing if the memory savings take place on the libMesh side or on the PETSc side, or on both sides.

Thanks for help.
Lorenzo



On May 30, 2013, at 4:23 PM, "Kirk, Benjamin (JSC-EG311)" <ben...@na...> wrote:

> On May 28, 2013, at 11:22 AM, lorenzo alessio botti <lor...@gm...> wrote:
> 
>> Yes you are right, it would be limited to dG, all variables of the same polynomial degree and no local P refinement.
>> I know that this is very specific but the memory savings are quite encouraging.
> 
> To properly test for this case will require looking at all elements on all processors, and assuring that (nvars*ndofs) is invariant.  Not that I am against it, but definitely some work to get it right.
> 
> To provide an estimate for how much memory savings this might produce in your problem, could you manually hardcode the block size for your problem at hand and see what happens, reporting the results?
> 
> I *think* that should work…
> 
> -Ben
> 
>