bugGNU Octave - Bugs: bug #54179, sparse/eigs.m test failure

 
 

bug #54179: sparse/eigs.m test failure

Submitted by:  Andrew Janke <apjanke>
Submitted on:  Sun 24 Jun 2018 09:17:31 PM UTC  
 
Category:  Test Suite Severity:  3 - Normal
Priority:  5 - Normal Item Group:  Unexpected Error
Status:  Fixed Assigned to:  None
Originator Name:  Open/Closed:  Closed
Release:  4.4.0 Operating System:  Any

Add a New Comment(Rich Markup)
   

You are not logged in

Please log in, so followups can be emailed to you.

 

( Jump to the original submission)

Wed 04 Jul 2018 02:08:37 PM UTC, comment #13:

Pushed the change here http://hg.savannah.gnu.org/hgweb/octave/rev/8925207dbc26. Marking as fixed.

Rik <rik5>
Project Administrator
Wed 04 Jul 2018 03:59:42 AM UTC, comment #12:

The patch works for me.

Dmitri.
--

Dmitri A. Sergatskov <dasergatskov>
Tue 03 Jul 2018 04:02:49 PM UTC, comment #11:

I've attached a patch (54179.cset). It can be tested with

@Andrew or @Dmitri: Could you verify this solves the problem.

I simply reduced the number of iterations from 5 to 4 on the tests that Dmitri indicated were failing. It doesn't upset my installation (all 182 tests pass), but I have arpack-3.3.0.

(file #44499)

Rik <rik5>
Project Administrator
Tue 03 Jul 2018 11:21:36 AM UTC, comment #10:

I have two tests failing:

and yes, changing opts.maxit to 4 in both cases allow them to pass.

Dmitri.
--

Dmitri A. Sergatskov <dasergatskov>
Tue 03 Jul 2018 07:02:35 AM UTC, comment #9:

@Rik: from comment #6, it seems that it is just a single test.
@Dmitri: do you confirm that if you decrease opts.maxit to 4, Fedora 28 + arpack 3.6.1 passes the test?

Marco Caliari <caliari>
Project Member
Mon 02 Jul 2018 04:38:52 PM UTC, comment #8:

@Marco: For the time being, is it just a single test that we need to lower the iteration count on?

If so, it seems worthwhile to make that change now and postpone making the the long term decision about testing for unconverged eigenvalues.

If this is not the case then I would be okay with removing these tests.

Rik <rik5>
Project Administrator
Mon 25 Jun 2018 02:54:12 PM UTC, comment #7:

Four iterations are enough to get at least two eigenvalues, while three are not. When no eigenvalue is found at the desired precision, arpack gives error -14.

There were a lot of {blas,arpack,platform}-dependent failing tests since the introduction of NaNs as placeholders for unconverged eigenvalues. Are we sure we want to keep these tests?

Marco Caliari <caliari>
Project Member
Mon 25 Jun 2018 09:54:59 AM UTC, comment #6:

Under Octave 4.4.0 built against arpack 3.6.1 on macOS 10.13.5:

If I take the following test in eigs.m at lines 1439 through 1446:

and decrease opts.maxit from 5 to 4, then the test passes.

If I decrease opts.maxit further to 3, I get a different error, `eigs: error -14 in dneupd`:

This seems to agree with your analysis.

I am no math expert, so I am unable to evaluate these numbers as to correctness.

Andrew Janke <apjanke>
Mon 25 Jun 2018 06:57:24 AM UTC, comment #5:

The result given with arpack-3.6.1 (original submission) is correct (you can try to increase opts.maxit and compare). If you look at the changes between 3.5.0 and 3.6.1 https://github.com/opencollab/arpack-ng/blob/master/CHANGES you see just one change to the algorithm (btw, by me) which could be responsible of a faster convergence od arpack. Dimitri confirmed the original result on a different platform, so I think we should change the test in order to reflect this improvement in arpack. What happens if you decrease opts.maxit to 4 or even smaller numbers?

Marco Caliari <caliari>
Project Member
Sun 24 Jun 2018 10:30:23 PM UTC, comment #4:

> It appears arpack 3.6.1 is the problem.


Confirmed. I rebuilt my Octave against arpack 3.5.0, and `test eigs` now passes.

Andrew Janke <apjanke>
Sun 24 Jun 2018 09:56:05 PM UTC, comment #3:

Not sure if I'm using DYLD_INSERT_LIBRARIES right, but using it to pull in OpenBLAS appears to have no effect:

Andrew Janke <apjanke>
Sun 24 Jun 2018 09:52:38 PM UTC, comment #2:

It appears arpack 3.6.1 is the problem.
I can reproduce this problem on Fedora 28 when octave
linked with arpack 3.6.1. No problem with 3.5.0 (the default Fedora's arpack):

Dmitri.
--

Dmitri A. Sergatskov <dasergatskov>
Sun 24 Jun 2018 09:37:32 PM UTC, comment #1:

I think the blas library (veclib) is primary suspect.
Could you try some other one? E.g. Openblas?
Perhaps you can override it at run time with
DYLD_INSERT_LIBRARIES

Dmitri.
--

Dmitri A. Sergatskov <dasergatskov>
Sun 24 Jun 2018 09:17:31 PM UTC, original submission:

In Octave 4.4.0 on macOS 10.13.5, the sparse/eigs.m test is failing for me. This is with arpack 3.6.1 installed. Build logs attached.

Looks related, but not identical, to:

Andrew Janke <apjanke>

 

(Note: upload size limit is set to 16384 kB, after insertion of the required escape characters.)

Attach Files:
   
   
Comment:
   

 

Depends on the following items: None found

Items that depend on this one: None found

 

Carbon-Copy List
  • -email is unavailable- added by rik5 (Posted a comment)
  • -email is unavailable- added by caliari (Posted a comment)
  • -email is unavailable- added by dasergatskov (Posted a comment)
  • -email is unavailable- added by apjanke (Submitted the item)
  •  

    Do you think this task is very important?
    If so, you can add your encouragement to it.
    This task has 0 encouragements so far.

    Only project members can vote.

     

     

     

    Follow 8 latest changes.

    Date Changed by Updated Field Previous Value => Replaced by
    2018-07-04 rik5 StatusIn Progress => Fixed
        Open/ClosedOpen => Closed
    2018-07-03 rik5 Attached File- => Added 54179.cset, #44499
    2018-06-25 caliari StatusNone => In Progress
    2018-06-25 caliari Item GroupIncorrect Result => Unexpected Error
        Operating SystemMac OS => Any
        Summarysparse/eigs.m test failure on macOS => sparse/eigs.m test failure
    2018-06-24 apjanke Attached File- => Added octave build logs - e7420cac8e489b4c38f287c587624cf3-a945c8e5c08b20b69682a6eaf08c5557f752991f.zip, #44425

    Back to the top


    Powered by Savane 3.3