[Q-e-developers] Diagonalization method as input flag in ph.x

[William Parker]

Paolo,

Could you outline how you would investigate a situation where CG works but Davidson fails?  Is it sensitive to ecutwfc? ecutrho? nbnd?  Not that I promise to investigate the issue extensively, but I will ask my colleague for the simplest case he knows of where he uses CG in ph.x, and what he tried.

--William

On Dec 21, 2012, at 11:55 AM, Paolo Giannozzi wrote:

> On Fri, 2012-12-21 at 11:45 -0600, William Parker wrote:
> 
>> I modified my ph.x to compare it and Davidson on my test cases, 
>> and a colleague of mine says that on a few systems he's had to 
>> switch to conjugate gradient to get convergence.  Is there any 
>> reason not to have it set as an input flag (default to 'david' 
>> but optional 'cg') like in pw.x?
> 
> no special reason, apart that it's yet another option.
> I would prefer that the reasons for occasional failures 
> of the Davidson algorithm would be elucidated and removed, 
> rather than resorting to a not-so-brilliant workaround 
> (I wrote the CG routine: I have the right to criticise it), 
> but I agree that no solution and a workaround is better 
> than no solution and no workaround
> 
> Paolo
> -- 
> Paolo Giannozzi, IOM-Democritos and University of Udine, Italy
> 
> 
> _______________________________________________
> Q-e-developers mailing list
> Q-e-developers at qe-forge.org
> http://qe-forge.org/mailman/listinfo/q-e-developers

*********************************************************
  William D. Parker              phone: +1 (630) 252-3388
  Argonne Scholar
  Argonne Leadership Computing Facility
  Building 240, Office 1-D.26
  Argonne National Laboratory
  9700 S. Cass Ave.
  Argonne, IL 60439
  U.S.A.
*********************************************************