[Scipy-tickets] [SciPy] #1472: Recent Arpack changes segfaulting in test suite

SciPy Trac scipy-tickets@scipy....
Sun Jul 3 10:36:23 CDT 2011


#1472: Recent Arpack changes segfaulting in test suite
---------------------------------+------------------------------------------
 Reporter:  rgommers             |       Owner:  pv    
     Type:  defect               |      Status:  new   
 Priority:  high                 |   Milestone:  0.10.0
Component:  scipy.sparse.linalg  |     Version:  0.9.0 
 Keywords:  arpack               |  
---------------------------------+------------------------------------------

Comment(by pv):

 The remaining failure with `-O0` is probably unrelated as it seems to be
 only a small precision issue. So it seems the crash is not in ARPACK, but
 in GMRES.

 Which of the tests in `isolve` triggers the crash [run
 "test(verbose=10)"]? Does it deal with complex numbers? Do the `isolve`
 tests pass, if you comment out complex-valued test cases? Also, can you
 try if you see the crash at commit:575262f3 -- the commit:e00f70c5
 following it introduced some fixes to the Fortran codes. You will see
 failures, but do you see a crash? (I do not see anything in
 commit:e00f70c5 that could cause a crash -- the changes are trivial.)

 I'm also surprised to see Valgrind crash like that -- it should instead
 print a report of an invalid read/write and that the program crashed.
 Maybe it doesn't work so well on OSX then.

 To get a more useful traceback, you should (i) set a breakpoint in the
 gmres routine (which may be named gmres_ or gmres__ etc. -- check objdump
 -T of the .so library) and follow the execution until it crashes, (ii) and
 write a simple Python snippet that induces a crash which only calls the
 iterative solver once, so that doing (i) becomes easier. The latter can be
 done by looking at which iteration of the test generator it crashes (the
 arguments are printed with verbose=10) in the test output. The former
 requires some experience with gdb -- it's difficult to give detailed
 instructions here.

-- 
Ticket URL: <http://projects.scipy.org/scipy/ticket/1472#comment:7>
SciPy <http://www.scipy.org>
SciPy is open-source software for mathematics, science, and engineering.


More information about the Scipy-tickets mailing list