[Trilinos-Users] AztecOO convergence

[Hoang Giang Bui]

Dear Dr. Mike

I used the same convergence criteria. I set the restart to maxIters but 
the convergence is not improved. I doubt that the preconditioner is not 
good enough to provide substantial speed up. Using gmres_condnum, I 
observed the estimated condition number ~10e11.

BR
Bui


On 06/28/13 23:00, Heroux, Michael A wrote:
> Bui,
>
> Just to be clear. The 1e-12 residual value comes from computing the 
> same residual expression?  In other words, with AztecOO you are 
> scaling the residual by the norm of the RHS.  Does the 1e-12 also 
> reflect scaling by the RHS?
>
> I just want to make sure you are doing an "apples-to-apples" 
> comparison of residuals.
>
> If the residual expressions are the same, the issue could be a lack of 
> convergence.  Are you restarting the GMRES iterations?  If so, try 
> increasing the restart value.  For instance, set it to 1000 and see if 
> convergence improves.  You may find that you will converge much 
> quicker and then be able to set the restart size to a smaller value. 
>  The value you want to set is AZ_kspace:
>
>   solver.SetAztecOption(AZ_kspace, maxIters);
>
>
> Where  maxIters is the maximum number of iterations you want to perform.
>
> In general, for tough problems, I try to avoid any restart at all.
>
> Let me know what you find out.
>
> Mike
>
> From: Hoang Giang Bui <hgbk2008 at gmail.com <mailto:hgbk2008 at gmail.com>>
> Date: Friday, June 28, 2013 7:57 AM
> To: "Ramsey, James J CIV (US)" <james.j.ramsey14.civ at mail.mil 
> <mailto:james.j.ramsey14.civ at mail.mil>>, 
> "trilinos-users at software.sandia.gov 
> <mailto:trilinos-users at software.sandia.gov>" 
> <trilinos-users at software.sandia.gov 
> <mailto:trilinos-users at software.sandia.gov>>
> Subject: [EXTERNAL] Re: [Trilinos-Users] AztecOO convergence
>
>
> That's correct. I used it to benchmark the iterative one.
>
> Bui
>
>
>
> On Fri, Jun 28, 2013 at 1:58 PM, Ramsey, James J CIV (US) 
> <james.j.ramsey14.civ at mail.mil <mailto:james.j.ramsey14.civ at mail.mil>> 
> wrote:
>
>     ----------------------------------------------------------------------------------------------------------------------------
>
>     The matrix is scaled in my problem. Before solving this, I do a
>     LeftScaling and observe that it gives better convergence. However,
>     considering that MKL solver can solve my problem with tolerance
>     1e-12. I'm confused that iterative solver cannot give the result
>     with the required tolerance.
>
>     ----------------------------------------------------------------------------------------------------------------------------
>
>     Is the MKL solver a direct solver?
>
>
>
>
>
>
>
>
>
> -- 
> With Best Regards !
> Giang Bui
> To learn and to excel

-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://software.sandia.gov/pipermail/trilinos-users/attachments/20130711/11df804a/attachment.html