[Trilinos-Users] MinTraceDavidson
Alicia Klinvex
aklinvex at purdue.edu
Fri Feb 27 07:55:36 MST 2015
Hello Sander,
I understand that you want some information about the inner iterations. I
need the timer output so that I can point that information out to you.
Does TraceMinDavidson print information about its running time when solve()
returns? If not, you need to enable that through your parameter list,
e.g. MyPL.set ("Verbosity", Anasazi::TimingDetails).
- Alicia
On Fri, Feb 27, 2015 at 9:45 AM, Sander Schaffner <ssander at student.ethz.ch>
wrote:
> Hey
>
> I'm just wondering how I can reach some information about the inner
> iteration (number of iterations, time to solution, ...), since the effect
> of the preconditioner should be seen there. Is there a debug flag I have to
> set?
>
> Best wishes,
> Sander
>
> Am 27.02.2015 um 15:38 schrieb Alicia Klinvex:
>
> Hello,
>
> Can you send me the output you're seeing of the timers? I'll help you
> interpret them.
>
> Ugh. "The number of operations Op*x" You just reminded me that it's
> inaccurate, because it does not include any of the matvecs from the inner
> Krylov iterations. Sorry about that. If you want the total number of
> Krylov iterations, we can also get that information from the timers.
>
> Best wishes,
> Alicia
>
> On Fri, Feb 27, 2015 at 9:15 AM, Sander Schaffner <ssander at student.ethz.ch
> > wrote:
>
>> Hey Alicia
>>
>> Thanks for the update on public Trilinos.
>>
>> I'm at point where I'm measuring the performance of my code. I there any
>> possibility to get the time of the inner interations (or did I miss this in
>> the output?)? And is the number of iteration steps in the inner loop
>> (solving the saddle point problem) indirectly given by "The number of
>> operations Op*x"? For example if it raises from one iteration step to the
>> next by 35 and I have a block size of 5, then it took 7 inner iteration
>> steps?
>>
>> Thanks
>>
>> Sander
>>
>> -------- Weitergeleitete Nachricht -------- Betreff: Re:
>> MinTraceDavidson Datum: Tue, 17 Feb 2015 17:09:18 +0100 Von: Sander
>> Schaffner <ssander at student.ethz.ch> <ssander at student.ethz.ch> An: Alicia
>> Klinvex <aklinvex at purdue.edu> <aklinvex at purdue.edu>
>>
>>
>> the jumping eigenvalue problem never occured again.
>>
>> I actually have two small questions:
>>
>> 1: If we run the solver with output I see the iterations step of the
>> outer loop as I understand it. So I dont see how it finds the new vectors
>> to enlarge the searchspace. Is there a way to turn that on?
>>
>> 2: So far I never got faster results if I activated a preconditioner. I
>> tried ifpack2:diagonal, ifpack2:ilut, muelu (smoother gauss, coarse klu2)
>> and a bandmatrix which I handed to ifpack2:ilut. Do you use preconditioners
>> and if so which?
>>
>> Thanks
>>
>> Sander
>>
>> Am 13.02.2015 um 18:44 schrieb Alicia Klinvex:
>>
>> It should find the three 0 eigenvalues first, and it does. They just
>> never get flagged for convergence if you monitor the RELATIVE residual.
>> Anything relative to 0 is going to be huge.
>>
>> The block preconditioning should probably hit the repo next week.
>>
>> Are you still seeing the problem you were before where small
>> eigenvalues jump in out of nowhere?
>>
>> - Alicia
>>
>> On Fri, Feb 13, 2015 at 7:57 AM, Sander Schaffner <
>> ssander at student.ethz.ch> wrote:
>>
>>> Hi Alica
>>>
>>> The matrix A that I have given you has a slightly other nullspace. Each
>>> value has to be corrected with the atom_mass.dat vector. Just have a look
>>> at the test_nullspace script i wrote today (also in the dropbox folder).
>>> Sorry for the confusion! The point is this: In the beginning I had to work
>>> on a matrix with no dimensions (physically) and we changed this for the new
>>> one. So each term in A is also corrected with the atom_mass. That's why we
>>> have another nullspace now (each 1 has to be multiplied by the square root
>>> of the according entry of atom_mass.dat)
>>>
>>> And I was also able to calculate many more eigenpairs. I'm now on 300 in
>>> 3 hours with ILUT as a preconditioner. And yes the smallest of your
>>> callculation is correct, altough I'm a bit confused since it should find
>>> the 3 eigenvalues 0 first?
>>>
>>> Is there any progress regarding other sattle point solver for the public
>>> repository of TraceMinDavidson?
>>>
>>> Best wishes,
>>> Sander
>>>
>>>
>>> Am 12.02.2015 um 20:34 schrieb Alicia Klinvex:
>>>
>>> Hello Sander,
>>>
>>> I'm copying the user list, since we probably should have been doing
>>> that all along. This way, you can get help from all the Trilinos folks,
>>> not just me. (I'm definitely the right person to ask about TraceMin, but I
>>> don't know much about Ifpack/ML, etc.) For everyone who is not Sander,
>>> here is a recap of his situation: He wants to compute the smallest
>>> eigenpairs of a symmetric positive definite matrix, and he's been brave
>>> enough to try Anasazi's newest eigensolver, TraceMin-Davidson. He is
>>> having trouble with TraceMin-Davidson, as well as a few of the other
>>> eigensolvers.
>>>
>>> I had a look at your matrix in Matlab, and it's a little unclear to me
>>> what the null space is. From your code, it looks like the basis is
>>> supposed to be [1 0 0 1 0 0 1 0 0 ...; 0 1 0 0 1 0 0 1 0 ...; 0 0 1 0 0 1 0
>>> 0 1]'. When I imported your matrix in Matlab and computed A*[your null
>>> space basis], I was getting a nonzero vector (and by nonzero, I mean its
>>> norm was 1.2685).
>>>
>>> I ran TraceMin-Davidson without calling setAuxVecs, and I did not
>>> experience the behavior you did. (150 eigenpairs were locked, but the
>>> three vectors of the null space never converged...which isn't surprising,
>>> considering that I asked for a small relative residual, and for those
>>> vectors, the relative residual is defined as absolute residual / 0.) Is it
>>> possible that your basis for the null space is incorrect, or did I misread
>>> your code?
>>>
>>> Best wishes,
>>> Alicia
>>>
>>> PS: The smallest eigenvalue TraceMin-Davidson found was 1.869423e-2.
>>> Is the matrix you sent me the same one you used to generate your results?
>>>
>>> On Wed, Feb 4, 2015 at 2:49 PM, Sander Schaffner <
>>> ssander at student.ethz.ch> wrote:
>>>
>>>> Of course: matrixFile.mtx -> its about 2gb
>>>>
>>>> Best wishes
>>>>
>>>> Sander
>>>>
>>>> Am 04.02.2015 um 20:21 schrieb Alicia Klinvex:
>>>>
>>>> Hello Sander,
>>>>
>>>> I don't see the matrices in your dropbox. Can you save them as
>>>> [something].mtx?
>>>>
>>>> Thank you,
>>>> Alicia
>>>>
>>>> On Wed, Feb 4, 2015 at 2:01 PM, Sander Schaffner <
>>>> ssander at student.ethz.ch> wrote:
>>>>
>>>>> Hey!
>>>>>
>>>>> It seems that it is not reproducable. I ran it again 2 times for 8
>>>>> hours and once I got 129 eigenpairs and the other time 230. Meanwhile the
>>>>> matrix is in the dropboxfolder in a MatrixMarket format. If you could feed
>>>>> this matrix once into your setup it would be nice.
>>>>>
>>>>> Thanks
>>>>>
>>>>> Sander
>>>>>
>>>>> Am 02.02.2015 um 16:18 schrieb Alicia Klinvex:
>>>>>
>>>>> Hello,
>>>>>
>>>>> Can you save your matrix as a MatrixMarket or Matlab file? Also,
>>>>> can you turn on the debugging output and rerun the program?
>>>>>
>>>>> Thank you,
>>>>> Alicia
>>>>>
>>>>> On Mon, Feb 2, 2015 at 6:21 AM, Sander Schaffner <
>>>>> ssander at student.ethz.ch> wrote:
>>>>>
>>>>>> Hey Alicia
>>>>>>
>>>>>> I hope this works for you:
>>>>>>
>>>>>>
>>>>>> https://www.dropbox.com/sh/ouw90m0m00s6kyn/AAAilijMstKYKa_7Mx4hQaQ9a?dl=0
>>>>>>
>>>>>> I had to write my own TpetraExt for the functions I need. Just copy
>>>>>> all the files in the same directory and change the trilinos-path in the
>>>>>> Makefile. In the sbatch file you see how I called the programm after
>>>>>> compilation.
>>>>>>
>>>>>> Is this okey for you? Thanks alot
>>>>>>
>>>>>> Best wishes,
>>>>>> Sander
>>>>>>
>>>>>>
>>>>>>
>>>>>> Am 30.01.2015 um 18:52 schrieb Alicia Klinvex:
>>>>>>
>>>>>> Well that's not good! It looks like there may be a problem with the
>>>>>> restarting. (I'll have a look at it.)
>>>>>>
>>>>>> My update isn't in the public repo yet, but thanks for reminding
>>>>>> me. I've been running tests with it, and if you're interested in using
>>>>>> preconditioners, the block diagonal preconditioning patch will be of
>>>>>> interest to you. (Mark Hoemmen generally helps me out with these things,
>>>>>> i.e. wrapping sensitive bits of code so I don't break the Trilinos
>>>>>> configuration process, but he was on vacation for most of January. I'll
>>>>>> ping him to see if he can do it now.)
>>>>>>
>>>>>> Can you send me your matrix and driver, so I can see what's going
>>>>>> on? I suspect the solver manager accidentally removed the wrong vector
>>>>>> from the subspace when it locked a vector.
>>>>>>
>>>>>> Best wishes,
>>>>>> Alicia
>>>>>>
>>>>>> PS: as far as getting unconverged Ritz vectors goes...it's
>>>>>> complicated. It's SUPPOSED to be difficult to access them, because they
>>>>>> aren't going to be accurate anyway. What you have to do if you desperately
>>>>>> want to get them is create your own custom debug status test. I can give
>>>>>> you more information about this if you like, but I don't think the
>>>>>> unconverged Ritz vectors will give you much information. You could always
>>>>>> turn on the Anasazi::Debug output option, which will give you information
>>>>>> about loss of orthogonality and whatnot.
>>>>>>
>>>>>> On Fri, Jan 30, 2015 at 9:51 AM, Sander Schaffner <
>>>>>> ssander at student.ethz.ch> wrote:
>>>>>>
>>>>>>> Hi Alicia
>>>>>>>
>>>>>>> There is a new strange thing that happened to me: I was able to
>>>>>>> calculate 400 eigenpairs of matrix A (dim = 150k, nnz = 60e9) in 4.5 hours.
>>>>>>> Then I received a updated matrix. The structure stayed the same, but the
>>>>>>> values got an update by either the square root of mass a or mass b.
>>>>>>> Therefore the values changed from 1e1 to 1e27. So I had to rescale, since
>>>>>>> the solver did not work for the values at 1e27. I have now a matrix B with
>>>>>>> the same structure than A but other values that have more or less the same
>>>>>>> size.
>>>>>>>
>>>>>>> I first calculated 100 eigenpairs which worked in 20min. But I did
>>>>>>> not give me any results for 150 eigenpairs after 8 hours. The solver doesnt
>>>>>>> converge. I had a look at the at the output. After finding 77 eigenpairs (I
>>>>>>> start with 3 auxiliary vectors) where the smallest eigenvalue is 1.92e-4 it
>>>>>>> suddenly has a new ritz-value of 1.2e-7 after iteration step 23 (it
>>>>>>> somehow looks af it whould smugling itself into the list). And this value
>>>>>>> must be wrong since we should not drop behind 1.92e-4. Do you have any idea
>>>>>>> how to correct this or why this happens? In the lower part of the mail you
>>>>>>> find two snippets of the TraceMinBase Solver Status. I thought I should
>>>>>>> have a look at the vector which belongs to this value, but I did not find
>>>>>>> how to configure the solver so that it stops after a certain number of
>>>>>>> iterations. Is there any possibility to get the ritz-vectors altough the
>>>>>>> solver did not finish?
>>>>>>>
>>>>>>> At the moment I'm using ILUT as a preconditioner. I'm working to get
>>>>>>> MueLu to run and use this to look if I can get any improvements.
>>>>>>>
>>>>>>> Then there is another question: Did your update made it into the
>>>>>>> public repo? I would love to try other saddle point solvers:)
>>>>>>>
>>>>>>> Best wishes,
>>>>>>>
>>>>>>> Sander
>>>>>>>
>>>>>>>
>>>>>>> ================================================================================
>>>>>>>
>>>>>>> TraceMinBase Solver Status
>>>>>>>
>>>>>>> The solver is initialized.
>>>>>>> The number of iterations performed is 23
>>>>>>> The block size is 50
>>>>>>> The number of blocks is 30
>>>>>>> The current basis size is 450
>>>>>>> The number of auxiliary vectors is 80
>>>>>>> The number of operations Op*x is 1200
>>>>>>> The number of operations M*x is 0
>>>>>>>
>>>>>>> CURRENT EIGENVALUE ESTIMATES
>>>>>>> Eigenvalue Residual(M) Residual(2)
>>>>>>>
>>>>>>> --------------------------------------------------------------------------------
>>>>>>> 5.476905e-04 not current 6.561567e-06
>>>>>>> 8.015696e-04 not current 6.822561e-06
>>>>>>> 8.067189e-04 not current 8.178570e-06
>>>>>>> 1.043989e-03 not current 2.537229e-05
>>>>>>> 1.063398e-03 not current 1.999694e-05
>>>>>>> 1.077780e-03 not current 2.646044e-05
>>>>>>> 1.148442e-03 not current 3.388691e-05
>>>>>>> 1.189898e-03 not current 1.205903e-04
>>>>>>>
>>>>>>>
>>>>>>> .
>>>>>>> .
>>>>>>> .
>>>>>>>
>>>>>>>
>>>>>>> ================================================================================
>>>>>>>
>>>>>>> TraceMinBase Solver Status
>>>>>>>
>>>>>>> The solver is initialized.
>>>>>>> The number of iterations performed is 24
>>>>>>> The block size is 50
>>>>>>> The number of blocks is 30
>>>>>>> The current basis size is 300
>>>>>>> The number of auxiliary vectors is 81
>>>>>>> The number of operations Op*x is 1250
>>>>>>> The number of operations M*x is 0
>>>>>>>
>>>>>>> CURRENT EIGENVALUE ESTIMATES
>>>>>>> Eigenvalue Residual(M) Residual(2)
>>>>>>>
>>>>>>> --------------------------------------------------------------------------------
>>>>>>> 1.266381e-07 not current not current
>>>>>>> 5.477028e-04 not current not current
>>>>>>> 8.015761e-04 not current not current
>>>>>>> 8.067281e-04 not current not current
>>>>>>> 1.044027e-03 not current not current
>>>>>>> 1.063419e-03 not current not current
>>>>>>> 1.077814e-03 not current not current
>>>>>>> 1.148475e-03 not current not current
>>>>>>> 1.190137e-03 not current not current
>>>>>>> 1.210246e-03 not current not current
>>>>>>> 1.230605e-03 not current not current
>>>>>>> 1.233197e-03 not current not current
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>
>>>
>>
>>
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://software.sandia.gov/pipermail/trilinos-users/attachments/20150227/00a710a3/attachment-0001.html>
More information about the Trilinos-Users
mailing list