[Libmesh-users] Re: [Libmesh-devel] Unused nodes

[Benjamin S. K. <be...@cf...>]

[I have sent this to the new libmesh-users list]

OK, I'll get you a patch that you can try.  It would be nice to have an 
option to remove the nodes, I agree...  This will certainly reduce the 
disk space required.

I'm still worried about the memory usage you report.  is the memory 
usage ~1/2 as much after 5000 time steps?  ~1/4 as much after 2500 time 
steps??  It seems like there could be a memory leak occuring at each 
time step.  Make sure you are running in debug mode & try the 
-log_summary option, which will print memory usage statistics (from 
PETSc) when the program quits.    The program valgrind can also be 
useful in finding memory leaks.

The only reason I mention this is because I routinely work with meshes 
that have on the order of 200k nodes and it doesn't require nearly that 
much memory.

I know what needs to be changed in addition to the patch you sent 
earlier.  I'll explain it more when I submit my patch later.

-Ben



Denis Danilov wrote:

>On Mon, Sep 22, 2003 at 11:40:00AM -0500, John Peterson wrote:
> 
>  
>
>>Yes this is a design decision.  During successive coarsening
>>and refinement, the same nodes could be created and destroyed
>>multiple times.  It was felt that keeping them in the nodes vector
>>was a reasonable tradeoff of space for performance.
>>    
>>
>
>This is right for quasi-stationary problems with weak time dependence,
>where only small number of cell will be coarsened (relative to number of
>refined cell). But what about "moving boundary" problems with high
>refined mesh only in the vicinity of moving interface? For example, when
>interface advances from one corner into the domain.
> 
>  
>
>>Is there a good reason you really need to do this?  For the patch
>>to be accepted, you would probably need to contribute a version
>>which can be turned on/off during configure similar to the infinite
>>elements.
>>    
>>
>
>There are one reason --- usage of memory and hard disk place. I run my
>program and see that it requires about 1.7 GB memory (after 10000 time
>steps, each time step includes mesh "refinement/coarsen"). For this time
>step the mesh consists of 36441 elements (27487 active) and 177099 nodes
>are stored in the mesh. I hope that memory usage will be reduced to ~700
>MB (or even less) if unused nodes will be deleted. The other thing is
>hard disk place. Look into out.gmv files of ex10 and you will see a
>lot of zeros in sections "nodes" and "variable". This is "solution" at
>unused nodes. So, in my case, with 36441 elements and 177099 the
>gmv-file consists almost fully of zeros.
>
>Regards,
>Denis
>
>
>
>-------------------------------------------------------
>This sf.net email is sponsored by:ThinkGeek
>Welcome to geek heaven.
>http://thinkgeek.com/sf
>_______________________________________________
>Libmesh-devel mailing list
>Lib...@li...
>https://lists.sourceforge.net/lists/listinfo/libmesh-devel
>  
>