On Oct 27, 2009, at 3:33 AM, Tim Kroeger wrote:

> I see.  Anyway, at least it should be possible to replace some of the
> pointers with ints or short ints (or perhaps chars).  I mean, one
> could make the arrays one-dimensional and then manage an array of
> offsets for the systems instead.  Since _n_v_comp and _dof_ids are
> private in DofObject, the necessary changes should be manageable.  Do
> you agree?  Would such a change be welcome if I implemented it?

This just feels like over optimization to me.  We're talking about  
_Mega_ bytes here.  I'm wary of making this more complicated / less  
understandable / slower to save MBs.  But, it's not really my call  
either...

> It would save some memory, but possible at the cost of time.  I don't
> know how often this information is accessed.  If it is read in each
> assembly, it could slow down things noticably.
>
> Still, there must be some other memory glutton in my application which
> I have not found yet.  Does anybody know how PETSc's memory
> consumption behaves in 64-bit mode?

Have you tried recompiling your application in 32-bit to make sure  
that's the problem?  Even on a 64bit machine you can compile for a  
32bit target and everything will run fine.  Have you checked out older  
versions of your application and libMesh to make sure it wasn't some  
change that you were unaware of?

I haven't noticed any large difference in memory usage in switching  
back and forth between 64bit and 32bit.  On my mac I used to regularly  
build 32bit with gcc and 64bit with Intel and never saw a 30%  
difference in the memory usage of my application.

I'm not saying it's not possible... but I do think you need to  
eliminate some variables before you drive yourself crazy chasing  
bytes...

Derek

------------------------------------------------------------------------------
Come build with us! The BlackBerry(R) Developer Conference in SF, CA
is the only developer event you need to attend this year. Jumpstart your
developing skills, take BlackBerry mobile applications to market and stay 
ahead of the curve. Join us from November 9 - 12, 2009. Register now!
http://p.sf.net/sfu/devconference
_______________________________________________
Libmesh-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/libmesh-users

Reply via email to