On 15-05-11 07:19 PM, Hong wrote:
Danyang:
I recently have some time-dependent cases that have difficulty in
convergence. It needs a lot of linear iterations during a specific
time, e.g., more than 100 linear iterations for every newton
iteration. In PETSc parallel version, this number will be doubled
or even more. Our case usually needs less than 10 iterations for
every newton iteration. This is NOT caused by PETSc relative
routines as the original sequential solver (ILU) also has the same
problem.
The results seems reasonable, but the large number of iterations
leads me to suspect that something is not well implemented in a
particular module. This module is developed a long time ago by
another guy and it sounds impossible for me to check every line
(thousands of lines) and do theoretical analysis.Does PETSc or
external package provide matrix analysis to help to locate the
'suspected' entry?
We need know what solvers being used. My guess is the default
gmres/bjacobi/ilu(0). Please run your code with option '-ts_view' or
'-snes_view' to find out.
Hi Hong,
Sorry for the late reply. I guess there is nothing wrong in the solver
or data structure, but maybe some large entries in the linear system
that cause the problem hard to solve. I will export the matrix to check
these values.
Thanks,
Danyang
Hong