[Trilinos-Users] Fwd: Krylov-Schur method with AztecOO
Heroux, Mike
MHeroux at CSBSJU.EDU
Mon Nov 17 10:44:51 MST 2014
Nelson,
AztecOO may balk at a zero diagonal value depending on the type of preconditioner you use. Generally speaking preconditioner implementations really prefer non-zero diagonals. I can’t tell from your output what preconditioner you are using, but if it is Jacobi, you will not converge, since Jacobi scales by diagonal values. If you are using an ILU type preconditioner, there is a possibility the preconditioner will work, if the zero diagonal values are not encountered prior to fill-in during the factorization. Even so, there is a risk that the factors will be ill-conditioned, leading to failure.
You could consider applying your own reordering of the matrix in order to get some term on all the diagonals. This is usually the best approach.
Mike
From: Alicia M Klinvex <aklinvex at purdue.edu<mailto:aklinvex at purdue.edu>>
Date: Monday, November 17, 2014 at 11:35 AM
To: "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>>, "maireni at cimne.upc.edu<mailto:maireni at cimne.upc.edu>" <maireni at cimne.upc.edu<mailto:maireni at cimne.upc.edu>>
Subject: [Trilinos-Users] Fwd: Krylov-Schur method with AztecOO
Hello,
I'm an external collaborator, so I really don't know much about AztecOO. (The only things I wrote in Trilinos are TraceMin and some of the things TraceMin depends on.) If you have questions about TraceMin, I would be happy to answer them. Otherwise, you're more likely to get a useful response by emailing the trilinos-users list. When you email the trilinos-users list, your message will reach ALL of the developers (including me).
Hopefully one of the AztecOO developers will be able to answer your question :-)
Best wishes,
Alicia
---------- Forwarded message ----------
From: <maireni at cimne.upc.edu<mailto:maireni at cimne.upc.edu>>
Date: Mon, Nov 17, 2014 at 11:11 AM
Subject: Re: Krylov-Schur method with AztecOO
To: Alicia Klinvex <aklinvex at purdue.edu<mailto:aklinvex at purdue.edu>>
Hi Alicia,
I understand how work Epetra matrix and i was able to
calculate eigen values with block Krylov-Schur method
using Amesos. However, changing to AztecOO i have a "nan"
as you see above. I dont know, but in this case
K and M are sparse stiffness and mass matrix respectively.
M is diagonal and have some zero terms in diagonal (rotational Dofs).
I don do understand what is the problem exactly.
Do you you have any idea about this mistake?.
Or just using AztecOO is limited to some special matrices,
due is a iterative solver.
Regards and best wishes
Nelson
iter: 0 residual = 1.000000e+00
iter: 1 residual = -nan
***************************************************************
Warning: a breakdown in this method
has occurred and solution has not converged.
Solver: bicgstab
number of iterations: 1
Recursive residual = -nan
Anasazi::BlockKrylovSchurSolMgr::solve() caught unexpected exception from Anasazi::BlockKrylovSchur::iterate() at iteration 1
/usr/local/trilinos-11.12.1/include/AnasaziEpetraAdapter.hpp:1401:
Throw number = 1
Throw test that evaluated to true: ret != 0
Anasazi::OperatorTraits<double,Epetra_Multivector,Epetra_Operator>::Apply(): Error in Epetra_Operator::Apply(). Code -2
Anasazi::BlockKrylovSchurSolMgr::solve() returning Unconverged with no solutions.
Anasazi::EigensolverMgr::solve() returned unconverged.
El 2014-11-13 02:49, Alicia Klinvex escribió:
The convergence rate of each eigenpair is based on the quantity
lambda_i / lambda_{s+1}, where s is the subspace dimension you
selected. TraceMin by default selects 2*the number of desired
eigenpairs. If the distribution of eigenvalues is bad, then it can
take many iterations. There are still things we can do to fix it,
like increasing the subspace dimension or using shifts.
Looking at that quantity lambda_i / lambda_{s+1} will tell you a lot
about why TraceMin is behaving this way.
On Tue, Nov 11, 2014 at 10:50 AM, <maireni at cimne.upc.edu<mailto:maireni at cimne.upc.edu>> wrote:
Hi,
I am using TraceMin for solve the problem.
My problem is symmetric and sparse.
However when i use lapack(is time consuming because is for dense
matrix)
FEAST or for example , TraceMin takes time for converge.
When i look for 4 eigen-values TraceMin is fast but when i increase
the number of eigen-values
to 20 TraceMin make a lot iteration. I was thinking that maybe i
have a wrong parameter.
I dont know.
Regards
Thanks
Nel
El 2014-11-11 23:15, Alicia M Klinvex escribió:
Hello,
I'm a little confused by the fact that you're using LAPACK and
TraceMin on the same problem. Is this a dense or sparse problem?
Yes, all of the Anasazi solvers should work with Tpetra as well as
Epetra.
Best wishes,
Alicia
----- Original Message -----
From: maireni at cimne.upc.edu<mailto:maireni at cimne.upc.edu>
To: "Alicia M Klinvex" <aklinvex at purdue.edu<mailto:aklinvex at purdue.edu>>
Sent: Tuesday, November 11, 2014 10:09:07 AM
Subject: Re: TraceMin-Davidson failurem
Hi Alice,
It is works now. I tested and give the similar values
of lapack eigen solver. However, why if increase the number
of eigen-values to looking for (at instance from 4 to 20
eigen-values),
it is take more time.
My toler is 1E-6.
I will try other eigen-solver too. I suppose Tpetra
can be use in the others eigen-solver.
Regards and thanks again
Nel
El 2014-11-11 01:51, Alicia M Klinvex escribió:
Hello,
Thank you so much for giving our newest solver a try!
Anasazi calls a function called isFillComplete, which tells us
whether
you finished putting things in a matrix. If you did NOT finish
putting things in the matrix, it breaks and outputs the error
message
you saw.
You may be thinking, "but I WAS finished modifying the entries of
the
matrix!" You need to let Anasazi know that by calling the Tpetra
function fillComplete
(http://trilinos.org/docs/dev/packages/tpetra/doc/html/classTpetra_1_1CrsMatrix.html#a27c06cd71c7542742a731a6cc91180d2
[1]).
In your code, this will look like K->fillComplete() and
M->fillComplete().
Does that answer your question?
Best wishes,
Alicia
----- Original Message -----
From: maireni at cimne.upc.edu<mailto:maireni at cimne.upc.edu>
To: "Alicia M Klinvex" <aklinvex at purdue.edu<mailto:aklinvex at purdue.edu>>
Sent: Monday, November 10, 2014 2:54:46 AM
Subject: Re: [Trilinos-Users] Returned mail: see transcript for
details
Hi Alicia,
At first so sorry for reply you a lot time for questions.
I am using the example you defined in the list that use
TraceMin-Davidson to solve a generalized eigenvalue problem
using Tpetra (and reading file).
The thing is i can not yet run my eigen problem.
In fact i have this mistakes and no reported in google
exist. Here it is.
Teuchos::ArrayRCP<unsigned
long>{ptr=0x46f9130,lowerOffset=0,upperOffset=8238,size=8239,node=0x4d6c7c0,strong_count=2,weak_count=0}
>> Allocating X_
>> Allocating KX_
>> Allocating MX_
>> Allocating R_
>> Allocating V_
>> Allocating KV_
>> Allocating MV_
>> done allocating.
Computing MV
terminate called after throwing an instance of 'std::runtime_error'
what():
/usr/local/trilinos-11.12.1/include/Tpetra_CrsMatrix_def.hpp:3517:
Throw number = 1
Throw test that evaluated to true: ! isFillComplete ()
What do you think i am doing wrong. I attached the code function
if you have time to check it for a while. I really interesting
about
you tell me before, because i want to use any eigen-solver you
defined
in anasazi.
Thanks a lot
Nelson
El 2014-11-06 22:20, Alicia M Klinvex escribió:
Hello,
a) The matrix should be the total matrix. If your matrix comes
from a
matrix-market or harwell-boeing file, we do have a function that
will
read the symmetric matrix from a file and store it as general. (I
believe this is to optimize matrix-vector multiplications.)
Alternatively, you can define your own matrix-vector multiplication
operator, in which case Anasazi never needs to see your matrix.
There's an example of this for TraceMin-Davidson.
b) Are you using BKS? If you're not using block Krylov-Schur, you
shouldn't need to create your own solver. RTR and
TraceMin/TraceMin-Davidson have solvers built-in.
c) I'm not sure I understand what you mean by the matrix being
dense
and sparse. Also, are you looking for the smallest or largest
eigenvalues? Are you looking for interior or extreme eigenvalues?
By the way, one of the cool things about Anasazi is that all of the
eigensolvers use the same interface. If you want to try a couple
of
them, it only involves changing maybe 5 lines of code in your
driver.
Best wishes,
Alicia
PS: We should copy the user list in our correspondence, since some
people are better equipped than I am to answer certain types of
questions. For instance, I am not an Amesos developer. The
Amesos
developers can probably do a better job answering your questions
about
Amesos.
----- Original Message -----
From: maireni at cimne.upc.edu<mailto:maireni at cimne.upc.edu>
To: "Alicia M Klinvex" <aklinvex at purdue.edu<mailto:aklinvex at purdue.edu>>
Sent: Thursday, November 6, 2014 7:05:34 AM
Subject: Re: [Trilinos-Users] Returned mail: see transcript for
details
Hello Alicia,
Just simple questions
a) When i use Anasazi with Tpetra in CSR format.
Should be the total matrix even if it is symmetric or it is ok if i
take
for Anasazi the upper or lower one?. How to do that?
b) I see Amesos accept now Pardiso solver. I use normally this
Solver.
My matrices is sparse and lapack is call when i use anasazi. How
can
i
set amesos for use Pardiso or SuperLU(shared Memory). I see some
macros
but, i think should be another way no?
c) The problem i have is symmetric dense and sparse. According
to your experience with is the fastest method in Anasazi?
Sorry again for ask you a lot, but i really excited with Trilinos.
Regards
Nelson
El 2014-11-06 10:38, Alicia M Klinvex escribió:
Hello,
Have you tried contacting the SLEPC team about your problems?
They're
a helpful group as well. (I'm obviously not trying to tell you
not
to
use Trilinos; I just want you to be aware of the packages that can
solve your specific problem. We don't currently support finding
all
eigenvalues in an interval. I should have that complete by the
end
of
the month though.)
FEAST is slow because of its reliance on direct solvers such as
PARDISO and all of the factorizations it requires. You may want
to
contact the FEAST developers to see whether it would be appropriate
to
reduce the number of contour integration points, which can speed up
the code.
Anasazi does have very powerful tools, and I'm glad you've decided
to
give them a try. :-)
I'm not familiar with AMLS. Are these small matrices dense or
sparse?
Regarding the linear systems, if they are sparse, Amesos and
Belos
are both really nice packages to use, depending on what your matrix
is
like. Amesos contains direct solvers, and Belos is for iterative
solvers such as CG.
Best wishes, and please keep us updated!
Alicia
----- Original Message -----
From: maireni at cimne.upc.edu<mailto:maireni at cimne.upc.edu>
To: "Alicia M Klinvex" <aklinvex at purdue.edu<mailto:aklinvex at purdue.edu>>
Sent: Wednesday, November 5, 2014 8:57:55 PM
Subject: Re: [Trilinos-Users] Returned mail: see transcript for
details
Hi Alicia M Klinvex,
Thanks for replying.
I was trying to use SLEPC's Krylov Schur but not successful yet.
I know the FEAST eigen-solver and is integrated in my code and
works.
I call directly from MKL libraries, but is slow than other
eigen-solvers,
at least block Lanzcos method like BLZPACK. Another problem is
ARPACK++,
that when i call the AReig functions not works correctly. It use a
format
of CSC. My matrices is symmetric and could be lower or upper
matrix.
In
any case,
the eigen-values is not ok. I saw Anasazi is inspired and have
better
tools than ARPACK. This is why i
try to use Anasazi.
Respect to size of problem, the idea is to use AMLS method. AMLS
method
is a condensation method
that can convert a huge matrix, millions of dofs, to just a little
thousands. In need to solve
a lot of eigen-values problems, but this eigen problems is so small
that
the corresponding one.
Also i need to solve a lot equations with a lot rigth hand side. I
was
thinking in use
Pardiso, but trilnos have good tools and i want to explot too.
Allow me try what you talk me and then i let you know.
Thanks for your help. Thanks a lot.
Nel
El 2014-11-06 01:27, Alicia M Klinvex escribió:
Hello Nelson,
What you want is this constructor:
http://trilinos.org/docs/r11.12/packages/tpetra/doc/html/classTpetra_1_1Map.html#aeb556362352e2ae62804601e36672a6c
[2]
Map( global_size_t numGlobalElements, GlobalOrdinal indexBase,
const
Teuchos::RCP< const Teuchos::Comm< int > > & comm, LocalGlobal lg =
GloballyDistributed, const Teuchos::RCP< Node > & node =
defaultArgNode< Node >() )
Here's what this constructor expects:
* The number of rows of your matrix
* The index base (which is generally 0)
* The communicator (which we will come back to)
* Whether the matrix is locally replicated or globally distributed
(which we will come back to)
* The node type, which defines the type of architecture you are
using
You may be thinking at this point, "I have to provide a
communicator?
But I don't want to use MPI!" You can define a serial Teuchos
communicator which never communicates, as in this example:
http://trilinos.org/docs/r11.12/packages/tpetra/doc/html/lesson01_no_mpi_8cpp-example.html
[3]
For a sequential example, it doesn't matter whether you specify
that
the matrix is locally replicated or globally distributed. We can
come
back to this when you're ready to use MPI.
Regarding finding all eigenvalues in an interval, you have several
options:
* You can compute the number of eigenvalues in an interval using
inertia computations (meaning sparse factorizations will be
involved),
then call any eigensolver you want.
* You can try Eric Polizzi's FEAST eigensolver (which requires an
estimate of the number of eigenvalues in an interval). You can
download it here: http://www.feast-solver.org/ [4] It can also be
found
in certain versions of the Intel Math Kernel Library.
* You can try SLEPC's Krylov Schur, which is supposed to support
this
feature. Just as a warning, I've personally never gotten it to
work
successfully.
We are currently working on adding that functionality (computing
all
eigenvalues in an interval) to Anasazi. Just out of curiosity,
roughly how many eigenvalues are there in your interval of
interest?
Also, how big is your matrix?
Best wishes,
Alicia
----- Original Message -----
From: maireni at cimne.upc.edu<mailto:maireni at cimne.upc.edu>
To: "Alicia M Klinvex" <aklinvex at purdue.edu<mailto:aklinvex at purdue.edu>>
Sent: Wednesday, November 5, 2014 1:08:00 AM
Subject: Re: [Trilinos-Users] Returned mail: see transcript for
details
Hi Alicia M Klinvex,
Thanks for reply.
In my case, as in research area, i compile Trilinos for
not running in MPI process, just in serial way. I know i can not
exploits
all tools of Trilinos, but is just for research. It have really
interesting tools.
When the code and the idea is working, migration to MPI is really
well.
For the case of column map and row map i need your help.
At the moment, just a class matrix works fine in Ananazi is ok for
me.
Epectra or Tpectra is ok. I am using now Tpectra and i see i can
define
a typedef
like this
typedef double scalar_type;
typedef Tpetra::CrsMatrix<scalar_type> TpetraCrsMatrix;
the other parameters is a default values, so i assumed with they
Anasazi works. My dude is in how to contruct the maps for serial
code.
Another thing i would like to ask you is about the eigen-solver.
Anasazi can provide me the number of desired eigenvalues,
but is possible it give to me an interval of eigen-values
i.e [0, fecuency cuttof]?
I really appreciate your help.
Best wishes,
Nelson
El 2014-11-04 21:57, Alicia M Klinvex escribió:
Hello Nelson,
My name is Alicia, and I am an Anasazi developer. I'm glad you've
decided to try Anasazi :-)
I assume what you're asking is how to take your three CSR arrays
and
turn them into a CrsMatrix that you can pass to an Anasazi
eigensolver. If you're using Tpetra, the constructor you're
looking
for is this:
http://trilinos.org/docs/r11.10/packages/tpetra/doc/html/classTpetra_1_1CrsMatrix.html#a139a91ad9430f012bfb3bcb8fe413ae1
[5]
(If you're using Epetra instead, or want to know the difference
between Epetra and Tpetra, let me know.)
Note that this constructor will expect the row map and column map,
which define how the data is distributed over the MPI processes.
If
you don't understand how to create those maps, please let me know
and
I'll help you out.
Does that answer your question?
Best wishes,
Alicia Klinvex
----- Original Message -----
From: maireni at cimne.upc.edu<mailto:maireni at cimne.upc.edu>
To: trilinos-users at software.sandia.gov<mailto:trilinos-users at software.sandia.gov>
Sent: Tuesday, November 4, 2014 3:41:49 AM
Subject: Re: [Trilinos-Users] Returned mail: see transcript for
details
Dear Heidi Thornquist,
Allow me introduce myself first. My name is Nelson Lafontaine,
i am a PhD student from UPC, Barcelona, Spain. Now i am Beijing
China
working with the Psor Chen Pu in the Automatic Multilevel
Substructure
Methods (AMLS).
I see you are a Anasazi developer and i write to you because i
am
interested in use
this library. My installation of Trilinos (trilinos-11.12.1) was
successfully,
even i can run some test examples that you are define in
Anasazi
files.
However, like you create your own class of matrices and vector,
my data i can provided to solve my eigen problem is in the form
of
a Compressed Sparse Row (CSR). My question is if exist a way
to
call
the eigen-solver defined in Anasazi just providing the data of
compressed
Sparse row format, or if is possible to create the matrices
types
defined
in Anasazi just using the CRS format.
Thanks in advance for your time.
Nelson
_______________________________________________
Trilinos-Users mailing list
Trilinos-Users at software.sandia.gov<mailto:Trilinos-Users at software.sandia.gov>
https://software.sandia.gov/mailman/listinfo/trilinos-users [6]
Links:
------
[1]
http://trilinos.org/docs/dev/packages/tpetra/doc/html/classTpetra_1_1CrsMatrix.html#a27c06cd71c7542742a731a6cc91180d2
[2]
http://trilinos.org/docs/r11.12/packages/tpetra/doc/html/classTpetra_1_1Map.html#aeb556362352e2ae62804601e36672a6c
[3]
http://trilinos.org/docs/r11.12/packages/tpetra/doc/html/lesson01_no_mpi_8cpp-example.html
[4] http://www.feast-solver.org/
[5]
http://trilinos.org/docs/r11.10/packages/tpetra/doc/html/classTpetra_1_1CrsMatrix.html#a139a91ad9430f012bfb3bcb8fe413ae1
[6] https://software.sandia.gov/mailman/listinfo/trilinos-users
More information about the Trilinos-Users
mailing list