Jed,
thanks for your reply.
By using the options you suggested, namely /-mg_levels_ksp_type
richardson -mg_levels_pc_type sor/, I was able to
solve without bumping into the DIVERGED_INDEFINITE_PC message.
Nevertheless, the number of iterations increases drastically as the
simulation progresses.
The Poisson's equation I am solving arises from a variable-density
projection method for incompressible multi-phase flows.
At each time step the system matrix coefficients change as a consequence
of the change in location of the heavier phase; the rhs changes
in time because of the change in the velocity field. Usually the
black-box multigrid or the deflated conjugate gradient method are used
to solve efficiently this type of problem: it is my understanding -
please correct me if I am wrong - that AMG is a generalization of the
former.
The only source term acting is gravity; the hydrostatic pressure is
removed from the governing equation in order to accommodate periodic
boundary conditions: this is more a hack than a clean solution. Could it
be the reason behind the poor performances/ DIVERGED_INDEFINITE_PC
problem I am experiencing?
Thanks,
Michele
On 05/17/2014 12:26 AM, Jed Brown wrote:
Michele Rosso <[email protected]> writes:
Hi,
I am solving an inhomogeneous Laplacian in 3D (basically a slightly
modified version of example ex34).
The laplacian is discretized by using a cell-center finite difference
7-point stencil with periodic BCs.
I am solving a time-dependent problem so the solution of the laplacian
is repeated at each time step with a different matrix (always SPD
though) and rhs. Also, the laplacian features large magnitude variations
in the coefficients. I solve by means of CG + GAMG as preconditioner.
Everything works fine for a while until I receive a
DIVERGED_INDEFINITE_PC message.
What is changing as you time step? Is there a nonlinearity that
activates suddenly? Especially a bifurcation or perhaps a source term
that is incompatible with the boundary conditions? You could try
-mg_levels_ksp_type richardson -mg_levels_pc_type sor. Can you
reproduce with a small problem?
The configuration looks okay to me.
Before checking my model is incorrect I would like to rule out the
possibility of improper use of the linear solver. I attached the full
output of a serial run with -log-summary -ksp_view
-ksp_converged_reason ksp_monitor_true_residual. I would appreciate if
you could help me in locating the issue.
