[Trilinos-Users] BlockKrylovSchur: number of required Ritz vectors splits a complex conjugate pair.
Michael Junge
junge at iam.uni-stuttgart.de
Thu Nov 27 23:45:45 MST 2008
Hi Heidi,
Yes you are right, a was using a modified version of the BKS in order to
track the residuals in each iteration step. I modified my code as
described, now it works just great!
Thanks a lot for your help!
Michael
Thornquist, Heidi K wrote:
> Hi Mike,
>
> Yes, it is easy to resolve this type of problem. The
> Anasazi::BlockKrylovSchur class is only an iteration, so this type of
> situation is normally handled by the solver manager class
> (Anasazi::BlockKrylovSchurSolMgr).
>
> If you look at lines 596-613 of the BlockKrylovSchurSolMgr.hpp file,
> you will see a block of code that gets the Ritz values and index from
> the BlockKrylovSchur class, determines whether a conjugate pair is on
> the boundary, and if necessary it modifies the number of Ritz pairs
> the BlockKrylovSchur class computes by calling “setNumRitzVectors”.
> From the error you received, it looks like you are working with a
> modified code because the original solver manager should have caught
> this problem and addressed it. So, you will want to have something
> like the block of code in the BlockKrylovSchurSolMgr.hpp file to
> address this situation.
>
> I hope this helps!
>
> Thanks,
> Heidi
>
>
>
> On 11/25/08 10:48 AM, "Michael Junge" <junge at iam.uni-stuttgart.de> wrote:
>
> Hi,
> when running the Block Krylov Schur solver of the trilinos package
> 8.0.3
> for an unsymmetric eigenvalue problem, I get the error message
> "Anasazi::BlockKrylovSchur::computeRitzVectors(): the number of
> required
> Ritz vectors splits a complex conjugate pair."
> after the 6-th iteration.
> Is there any possibility to cope with this error? Can I instead of
> throwing this exception
> lower the variable numRitzVecs_ by one and proceed without any
> subsequent error?
> Any suggestions to resolve this error are greatly appreciated.
>
> Michael
>
>
>
More information about the Trilinos-Users
mailing list