On Tue, Oct 30, 2012 at 10:07 AM, Roy Stogner <[email protected]>wrote:

> Interesting; I hadn't noticed that method when it went in.  Looks like
> you're out of luck if a single EquationSystems has more than one
> System with the same variable name, but then I suspect pretty much all
> our I/O other than xda/xdr fails in one way or another in that case.
>

Yeah - I've always assumed that that use case is just not supported.  Our
software simply doesn't allow it at all.


> Any complaints about that method's semantics?  If not, then when we
> ever get around to implementing that capability in other output
> formats it seems like the thing to do would be simply to move that
> method up from ExodusII_IO to MeshOutput and then examine the result
> in the other subclasses.


This would be really great.  For one thing... what I did in the Exodus
writer is sub-par because we still do all the parallel communication and
projection (if needed) for that variable... and then it just gets filtered
out at the Exodus writing stage.  That's still nice (because we have
simulations with thousands of variables... but typically only need to write
out a few) but it would be great to get rid of the extra overhead.

Oh - while we're on this subject.... I'll bring up again that
EquationSystems::build_solution_vector() is doing several full
serializations (all-to-all full copies) of the global solution vector.
 This is happening in System::update_global_solution() ....

When we're running rather large problems (100M DoFs+ on 10k+ processors)
this actually can be a significant issue: a full copy of the solution
vector for 100M+ DoFs is ~1GB... which, when you have 32 MPI processes on
the same box can cause quite a big issue!  Not too mention the time
necessary to do a full localization to every processor.

When we're using a parallel IO format (like Nemesis) we REALLY need to NOT
do this full localization.  The problem is that this is down in
EquationSystems and System (and quite complicated).  I've tried to fix it
before... but haven't really come up with great solution....

Derek
------------------------------------------------------------------------------
Everyone hates slow websites. So do we.
Make your web apps faster with AppDynamics
Download AppDynamics Lite for free today:
http://p.sf.net/sfu/appdyn_sfd2d_oct
_______________________________________________
Libmesh-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/libmesh-users

Reply via email to