[petsc-dev] strangness in Chebyshev estimate of eigenvalues

[Barry Smith]

> On May 17, 2016, at 1:18 PM, Jed Brown <jed at jedbrown.org> wrote:
> 
> Barry Smith <bsmith at mcs.anl.gov> writes:
>>  Of course, as always, with different number of processes one will
>>  get different results, even with Jacobi + Chebyshev because the
>>  random numbers generated will be different with different number of
>>  processes, is this important?
> 
> I can't think of a way to get a "noisy" vector deterministically and
> independent of the distribution.  We could use a hash of the global row
> index, which would help with debugging in the sense that we could read
> in a stored matrix on a different number of cores and get the same
> Chebyshev/Jacobi estimates.  I have a minor preference for this instead
> of random initialization, but it wouldn't help with most discretizations
> because the distribution usually changes the ordering.

  Mark,

  Sounds like Jed is saying he is ok with making the default use the random number generator. Do this in a branch and be aware that you have to update a bunch of the test output files to match the new convergence histories introduced by the change.

   Barry