[petsc-users] The same code randomly produce different residual norms

[Jed Brown]

Fande Kong <fd.kong at siat.ac.cn> writes:
> * mpirun  -n 8  ./ex5 -pc_type mg \ -ksp_monitor\ -pc_mg_levels 3
> \ -pc_mg_galerkin \ -da_grid_x 17 \ -da_grid_y 17
> \ -mg_levels_ksp_norm_type unpreconditioned \ -snes_monitor
> \ -mg_levels_ksp_chebyshev_estimate_eigenvalues 0.5,1.1
> \ -mg_levels_pc_type sor \ -pc_mg_type \*
>
>
> I run this script several times, but at each time, the residual norms have
> some tiny differences ( these should not happen). For example:

Some messages are unpacked in the order in which they are received,
rather than doing a moderately expensive sorting.  There are some
options to make VecScatter and other components reproducible, but this
configuration might use a feature for which such an option has not been
implemented (-vecscatter_reproduce helps a little in my test, but is not
sufficient).

The difference you're seeing is harmless, but if you have an application
for which this is required (either to reproduce rare events for
re-analysis of an unstable dynamical system or due to programmatic
requirements that misunderstand error analysis of finite-precision
arithmetic), we can hunt down any such places and provide an option to
make it reproducible.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 835 bytes
Desc: not available
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20140321/ba77832b/attachment.pgp>