[Trilinos-Users] Applying Anasazi::BlockKrylovSchurSolMgr

[Alicia Klinvex]

Hello Joseph,

Sorry for the delay.  The problem is that Block Krylov Schur can't solve
generalized eigenvalue problems without a spectral transformation.  For
instance, instead of solving Ax=\lambda Mx, you could ask it to solve
A^{-1} M x = 1/\lambda x or M^{-1} A x = \lambda x.  We have a few examples
of how to do this on the Anasazi page.

Please try a spectral transformation and see if that resolves your issue.

Best wishes, and happy holidays!
Alicia

On Mon, Dec 14, 2015 at 7:49 AM, Rutherford, Joseph M <jmruther at illinois.edu
> wrote:

> Alicia et al,
>
>
>
> Thanks for your very helpful reply.  I extended the size of my 4x4 system
> to 10x10 by increasing the dimension size in the serial Tpetra::Map
> instance I’m using for both domain and range. So now I still have a pair of
> rank 4 matrices, and the A x = \lambda B x system still should have the
> same eigenvalues and the eigenvectors should just have zero padding.
>
>
>
> The good news is that your suggestion worked very well; the system is now
> satisfied to move into the solve stage. The bad news is that the solver
> seems to be getting stuck in a loop in the first iteration.  Since I can
> SVD a dense random 100x100 system nearly instantaneously with NumPy, I
> think I must still be using BKS incorrectly if its first iteration for
> 10x10 is still underway after several minutes of full CPU usage.  Do you
> have any suggestions for how to dig further into this?  Debugging indicates
> I’m stuck in the while(doGramSchmidt){…} loop of int
> SVQBOrthoManager<ScalarType, MV, OP>::findBasis(…).  Terminal output below.
>
>
>
> Thanks again for getting me past that last hurdle.
>
>
>
> Joe
>
>
>
> <terminal>
>
> $ ./bin/test_eigensolver.exe
>
> Running 1 test case...
>
> Debugging checks: iteration 0: after initialize()
>
> >> Error in F^H M F == I  : 1.11e+01
>
>
>
>
> ================================================================================
>
>
>
>                          BlockKrylovSchur Solver Status
>
>
>
> The solver is initialized.
>
> The number of iterations performed is 0
>
> The block size is         1
>
> The number of blocks is   5
>
> The current basis size is 0
>
> The number of auxiliary vectors is 0
>
> The number of operations Op*x   is 0
>
>
>
> CURRENT RITZ VALUES
>
>    [ NONE COMPUTED ]
>
>
>
>
> ================================================================================
>
>
>
> - StatusTestOutput: Failed
>
>   (Num calls,Mod test,State test): (1, 1, Passed Failed Undefined )
>
>    - StatusTestCombo: Failed
>
>      - StatusTestWithOrdering: Undefined
>
>        Quorum: 4
>
>        Auxiliary values: [empty]
>
>        - StatusTestResNorm: Failed
>
>          (Tolerance,WhichNorm,Scaled,Quorum):
> (1.000000e-03,RITZRES_2NORM,true,4)
>
>          Which vectors: [empty]
>
> </terminal>
>
>
>
>
>
> *From:* Alicia Klinvex [mailto:aklinvex at purdue.edu]
> *Sent:* Sunday, December 13, 2015 12:33 AM
> *To:* Rutherford, Joseph M <jmruther at illinois.edu>
> *Cc:* trilinos-users at trilinos.org
> *Subject:* Re: [Trilinos-Users] Applying Anasazi::BlockKrylovSchurSolMgr
>
>
>
> Hello Joseph,
>
>
>
> BKS tries to build a subspace of dimension block_size * num_blocks.  If
> the dimension of your problem n is less than that subspace dimension, it's
> impossible to orthogonalize the basis vectors (since you can't generate 10
> orthogonal vectors within a subspace of dimension 4).  Perhaps we should
> modify the error message to make this clearer.
>
>
>
> Can you try a slightly larger example, perhaps 10x10, with block_size 1
> and num_blocks < 10?
>
>
>
> Best wishes,
>
> Alicia
>
>
>
> On Sat, Dec 12, 2015 at 1:57 PM, Rutherford, Joseph M <
> jmruther at illinois.edu> wrote:
>
> I’m having some difficulty using Anasazi to compute the smallest
> eigenvalues of a (non-Hermitian) complex-valued linear system.  Before
> addressing a general case,  I have written a simple test using two
> Tpetra::CrsMatrix<std::complex<float>,int,int> instances with reference
> data – this is a very small 4x4 dense case with reference data available.
> (I can share the example code if needed.)
>
>
>
> Unfortunately, I am having a hard time getting the SolMgr to drive the
> solution.  As I’m solving a reference 4x4 system with 4 nondegenerate
> eigenvalues, I’m fairly confident I don’t have a pathological case,
> although it is much smaller than the usual Anasazi use case.  Could you
> please look at the error message reported to my console (below)?  I’ve not
> found a clearly-aligned example in the Anasazi distribution, so I’m hopeful
> this is just a trivial case that is easily fixed.
>
>
>
> Sincerely yours,
>
>
>
> Joe Rutherford
>
>
>
> <console>
>
> $ ./bin/test_eigensolver.exe
>
> Running 1 test case...
>
> unknown location(0): fatal error: in "eigensolver_dense_4x4_complex":
> std::invalid_argument:
> /home/jmruther/branches/bean/externals/trilinos/include/AnasaziBlockKrylovSchurSolMgr.hpp:344:
>
>
>
> Throw number = 2
>
>
>
> Throw test that evaluated to true:
> static_cast<ptrdiff_t>(_numBlocks)*_blockSize >
> MVT::GetGlobalLength(*_problem->getInitVec())
>
>
>
> Anasazi::BlockKrylovSchurSolMgr: Potentially impossible orthogonality
> requests. Reduce basis size.
>
> /home/jmruther/branches/bean/source/tests/eigensolver.cpp(57): last
> checkpoint: "eigensolver_dense_4x4_complex" entry.
>
>
>
> *** 1 failure is detected in the test module "Eigensolver"
>
> $
>
> </console>
>
>
>
>
>
>
> _______________________________________________
> Trilinos-Users mailing list
> Trilinos-Users at trilinos.org
> https://trilinos.org/mailman/listinfo/trilinos-users
> <https://urldefense.proofpoint.com/v2/url?u=https-3A__trilinos.org_mailman_listinfo_trilinos-2Dusers&d=BQMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=Ng9lEtA9hn_MbIhqAtV5991RoWo8JGET5un-J3-Ub7o&m=a92OB0Je5c97qWnvqkK3eKwfHg65YXj6gsz4PO4b8P0&s=-Aqn6cS4l23CGmblqMY86E-cVM9F7hK4MpfCFtvnIkA&e=>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://trilinos.org/pipermail/trilinos-users/attachments/20151225/52a7acc0/attachment.html>