libMesh: A C++ Finite Element Library

Email Archive: libmesh-devel (read-only)

[Libmesh-devel] Changing Partitioners to take element weights.

From: Derek Gaston <derek.gaston@in...> - 2011-01-19 16:23

So, I wanted to be able to give some input into Metis in the way of element weights.  By default libMesh just weights each element with it's number of nodes.  What I added gives you the ability to attach an ErrorVector * that is then used to weight the partitioning.  This is working well for my purposes.

Why an ErrorVector?  Because I thought that having some of the statistics capabilities of ErrorVector might be useful to someone doing partition weighting (things like cut_above, variance, mean, etc.).  I am already using some of those capabilities in my own code.

I'm attaching a diff that makes these changes just to MetisPartitioner.  I wanted to see if you guys are ok with this change.... or if I should actually make some changes to the base Partitioner class to allow this.  I honestly don't know what the right thing to do is.  I would hate to have someone attach weights only to have them ignored by something like the LinearPartitioner or SFCPartitioner.

What do you think?

Derek

Thread View

Thread	Author	Date
[Libmesh-devel] Changing Partitioners to take element weights.	Derek Gaston <derek.gaston@in...>

Re: [Libmesh-devel] Changing Partitioners to take element weights.

From: Roy Stogner <roystgnr@ic...> - 2011-01-19 16:50

On Wed, 19 Jan 2011, Derek Gaston wrote:

> What do you think?

Your patch looks good.  I agree with the choice of ErrorVector (with
the caveat that this makes it even more important for us to properly
parallelize ErrorVector...).  I'd put attach_weights() in the base
class with a default libmesh_not_implemented() implementation.
---
Roy

Re: [Libmesh-devel] Changing Partitioners to take element weights.

From: Derek Gaston <derek.gaston@in...> - 2011-01-19 16:55

On Jan 19, 2011, at 9:50 AM, Roy Stogner wrote:

> Your patch looks good.  I agree with the choice of ErrorVector (with
> the caveat that this makes it even more important for us to properly
> parallelize ErrorVector...).

Yeah - there is actually some weirdness when trying to do Parallel::sum(ErrorVector).  I ended up static casting it to a std::vector<ErrorVectorReal> before passing it into Parallel::sum().  I don't have time to look into it now... but we should probably take care of that at some point in the future.

> I'd put attach_weights() in the base
> class with a default libmesh_not_implemented() implementation.

Ok - sounds good.  I'll do that and check it in.

Derek