[Trilinos-Users] [EXTERNAL] Anasazi LI and SI bug + patch

Thornquist, Heidi K hkthorn at sandia.gov
Tue Sep 1 13:34:58 EDT 2015


Hi Pavel,

Yes, you are correct about that bug.  I will look over the patch and apply it as soon as possible.

Thanks,
Heidi

--

Heidi K. Thornquist
Electrical Models & Simulation
Sandia National Laboratories
Albuquerque, NM  87185-1323


From: Trilinos-Users <trilinos-users-bounces at trilinos.org<mailto:trilinos-users-bounces at trilinos.org>> on behalf of Pavel Holoborodko <pavel at holoborodko.com<mailto:pavel at holoborodko.com>>
Date: Monday, August 31, 2015 at 11:27 PM
To: "trilinos-users at trilinos.org<mailto:trilinos-users at trilinos.org>" <trilinos-users at trilinos.org<mailto:trilinos-users at trilinos.org>>, "Hoemmen, Mark" <mhoemme at sandia.gov<mailto:mhoemme at sandia.gov>>
Subject: [EXTERNAL] [Trilinos-Users] Anasazi LI and SI bug + patch

Symptoms
BlockKrylovSchurSolMgr crashes when used in 'LI' and 'SI' modes.
(or TREXC ends with error "** on entry DTREXC, parameter number 7 had an illegal value")


Reason
BlockKrylovSchur requires that conjugate Ritz values are stored one after another.
(e.g. TREXC, sortRitzValues, etc.)

But, in case of 'LI' and 'SI', BasicSort::sort breaks conjugate pairs and might place them far apart after sorting.

As a result, this code is unpredictable, AnasaziBlockKrylovSchur.hpp, 1604-1605 lines:

    // Sort using both the real and imaginary parts of the Ritz values.
    sm_->sort(tmp_rRitzValues, tmp_iRitzValues, Teuchos::rcpFromRef(ritzOrder_) , curDim_); // <- places conjugate values apart
    HelperTraits<ScalarType>::sortRitzValues( tmp_rRitzValues, tmp_iRitzValues, &ritzValues_, &ritzOrder_, &ritzIndex_ ); // <- requires clustered conjugate pairs

Fix
Patch is in attachment.
It makes the BasicSort class to preserve clustered storage of conjugate eigenvalues in 'LI' and 'SI' modes.

***
Please consider applying the patch as now Anasazi is unable to operate in LI and SI modes (finding eigenvalues with largest/smallest imaginary parts).

Thank you,
Pavel.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://trilinos.org/pipermail/trilinos-users/attachments/20150901/8afe02b1/attachment.html>


More information about the Trilinos-Users mailing list