[petsc-users] Help with ML/BoomerAMG

[Jed Brown]

Gaetan Kenway <gaetank at gmail.com> writes:

> It is an SA turbulence model and the discrete adjoint computed exactly with
> AD. Certainly the grids are highly stretched in the BL since the grids are
> resolving the viscous sublayer (y+ < 1) and the  Reynolds numbers are on
> the order of  10's of millions. I tend  only to see this behaviour at
> higher mach numbers when stronger shocks start to appear.  For example, the
> adjoint  system may solve  fine at M=0.80, and fail to converge at M=0.85.

How meaningful is the information provided by the discrete adjoint here?
Limiters and even just upwind discretizations on non-uniform grids lead
to inconsistent discretizations of the adjoint equations.  If the
adjoint equation is full of numerical artifacts, it can cause the linear
problem to lose structure, resulting in singular sub-problems, negative
pivots, and other badness.  What happens when you use a direct solve for
subdomain problems (ASM+LU; use smaller subdomains if necessary)?