On Nov 11, 2011, at 11:20 AM, Jed Brown wrote:

> On Fri, Nov 11, 2011 at 10:08, Mark F. Adams <mark.adams at columbia.edu> 
> wrote:
> First, "my" matrices (actually XGC1 matrices from one of my projects and I 
> don't even know exactly what they do) put 1.0 on the diagonal for BCs and god 
> knows how the thing is scaled (1e13 apparently).  But this does not matter 
> because the RHS and initial guess are 0.0 so the BCs have been completely 
> removed from the algebra, even if they are still in the data structures.
> 
> So what happens when you use the same method with inhomogeneous Dirichlet 
> conditions?

Jed, I agreed with this point in my email.  And I can now see your point that 
using the preconditioned residual is a good default for PETSc, or at least I'm 
willing to defer to your judgment.

> This is especially bad if the interior is scaled by, say, 1e-13 instead of 
> 1e+13, because the boundary conditions dominate the initial residual.
>  
> 
> As I recall the preconditioned residual was confusing because since these are 
> Laplacian matrices with a scale of 1e13, as you found, the residual dropped 
> like 10 orders of magnitude in the first iteration, which was  pretty 
> confusing.
> 
> The preconditioned residual fixes the scaling. If you evaluate the initial 
> unpreconditioned residual, then you see the confusing scaling. But 
> preconditioning, even with just Jacobi, fixes the scaling.

All I recall was that I was confused because the residual dropped 10 orders of 
magnitude, apparently, in the first iteration and Barry explained the issue, I 
added 'unpreconditioned' residual and it was fine.  So if I am recalling this 
correctly this default confused me but apparently, or what I infer from this 
discussion, more users get confused with using the unpreconditioned residual so 
you've chosen preconditioned residual as the default.

Mark

>  
> 
> So I don't see these XGC1 problems as being arguments for preconditioned 
> residual, in fact they argue against it, right?
> 
> The preconditioned residual fixes the scaling, assuming it is no worse than 
> Jacobi.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: 
<http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20111111/e23023ed/attachment.html>

Reply via email to