Hi,
That’s an interesting observation — not sure I can claim that I’ve seen the same behaviour, but occasionally the iterative solver has been a bit unstable (particularly when e.g. using random noise to start up a transitional problem in the quasi-3D case).
There are a few things you might try:
- drop the solver tolerance. The default is 1e-9, but maybe something is going wrong with the normalisation we use on the pressure field which is making the pressure field converge to the wrong values
- try another solver type. Another option in parallel is XxtStaticCond or XxtMultiLevelStaticCond, both of which are parallel direct solvers and avoid iterative solvers completely. These could be quite a bit slower, depending on the size of your mesh.
- alternatively if your mesh isn’t too big, you can set the ‘npz’ command-line option to the same number of processors as your parallel simulation. In this case each plane will lie on a single processor, and you can then use a serial direct solver such as DirectMultiLevelStaticCond.
- If none of these solve the problem, I’d also advise looking at the mesh/boundary conditions/etc to make sure nothing has been set incorrectly.
Hope this helps.
Dear all,
I am running the timestepper of the IncNavierStokesSolver for
a Quasi-3D problem in parallel. As I mentioned in a
previous email, the default combination of linear solver and
preconditioning (allegedly IterativeStaticCond) seems to be
producing non-physical oscillations in the pressure field.
I have tried to follow the suggestion that non-default settings (and
different for velocity and pressure fields) like those specified in
the user guide as an example might improve the situation:
<GLOBALSYSSOLNINFO>
<V VAR="u,v,w">
<I PROPERTY="GlobalSysSoln"
VALUE="IterativeStaticCond" />
<I PROPERTY="Preconditioner"
VALUE="LowEnergyBlock"/>
<I PROPERTY="IterativeSolverTolerance"
VALUE="1e-8"/>
</V>
<V VAR="p">
<I PROPERTY="GlobalSysSoln"
VALUE="IterativeStaticCond" />
<I PROPERTY="Preconditioner"
VALUE="FullLinearSpaceWithLowEnergyBlock"/>
<I PROPERTY="IterativeSolverTolerance" VALUE="1e-6"/>
</V>
</GLOBALSYSSOLNINFO>
These preconditioners seem incompatible with Quasi-3D
discretisations and the solver crashes with an error. I have tried
different combinations but the only that seems to be working is Diagonal
for "u,v,w" and FullLinearSpaceWithDiagonal for "p".
Are these the defaults when a mere IterativeStaticCond is specified
in SOLVERINFO or nothing is said at all? If so, what are the default
values for IterativeSolverTolerance? Are these tolerances
the same for velocity and pressure fields? Is there any other
preconditioning option that I can use with quasi-3D problems?
Cheers