[petsc-users] Large Matrix eigenvalue problem : configuration

[Barry Smith]

> On May 18, 2015, at 5:07 PM, Alp Kalpalp <alpkalpalp at gmail.com> wrote:
> 
> Hi,
> 
> While reading the link you gave, I have seen a strange sentence;
> 
> "Otherwise you will get strange crashes."
> 
> can someone explain it why?

   Integer overflow is not really handled well in software. So when you add two int numbers together (each less than the maximum value that can go into an int, but whose sum won't fit into an int) you can end up with a negative number or a much smaller positive number. The code just keeps on chugging along with these "incorrect" numbers (producing incorrect results but not necessarily crashing or providing any indication to the user that something is wrong). Eventually if you are lucky something gets so out of whack it crashes and so you know something is wrong. If you are unlucky it never crashes and you think the results are fine when they are not.

  Barry



> 
> On Sun, May 17, 2015 at 12:16 PM, Jose E. Roman <jroman at dsic.upv.es> wrote:
> http://www.mcs.anl.gov/petsc/documentation/faq.html#with-64-bit-indices
> 
> 
> 
> El 17/05/2015, a las 09:02, venkatesh g escribió:
> 
> > Thank you. So the single eigenvector's size which is Nx1 is the number of unknowns and should be less than 2^31-1 so as to avoid configuring with '--with-64-bit-indices=1' ??
> >
> >
> > On Sat, May 16, 2015 at 8:05 PM, Matthew Knepley <knepley at gmail.com> wrote:
> > On Sat, May 16, 2015 at 3:49 AM, venkatesh g <venkateshgk.j at gmail.com> wrote:
> > Hi,
> >
> > I am trying to solving AX=lambda BX
> >
> > A and B are of size 57000 x 57000
> >
> > Currently I solve my small matrices using -st_type sinvert and -eps_nev 1
> >
> > So I get 1 eigenvalue and eigenvector.
> >
> > So for the 57000x57000 large matrix problem, the unknown eigenvector is 57000x1
> >
> > Should I configure the petsc as '--with-64-bit-indices=1' for this problem ?
> >
> > It is not necessary.
> >
> >    Matt
> >
> > Is the unknowns being considered as the entire eigenspace 57000x57000 ?
> >
> > kindly let me know.
> >
> > Venkatesh
> >
> >
> >
> >
> > --
> > What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
> > -- Norbert Wiener
> >
> 
>