Currently we run into issues with the parallel AMG if
redistribution happened on the coarsest level. That is
not detected by AMG and it will construct smoothers on
all processors present before the redistribution. Some of
them will get OwnerOverlapCommunication objects that have
an invalid MPI communicator in them (MPI_COMM_NULL) and an
MPI_ERROR will be raised as we communicate in the constructor.
With this patch we detect this situaton and set the pointer to
OwnerOverlapCommunication to null to prevent communication. A
sanity check that the matrix has zero rows has been added.
Often it is claimed by CPR-AMG evangelists that this will make the pressure
system more elliptic. That may be the case. But even more important it also
decouples the pressure from the saturations.
Without this approach the pressure correction influences the smoothing of the
full system too much and e.g. for Norne CPR is worse than simple block ILU0.
This seems to have been forgotten previously. Now the code int CPRPreconditioner.hpp
uses ParallelOverlappingILU0 instead of SeqILU[0n]/BlockPreconditioner which
makes the code more slim.
The approach is inspired by Geiger's system-amg but we use dune-istl
aggregation AMG for it. On the fine level all unknowns attached to a cell
form a matrix block and are treated fully coupled. To form the first
coarse level system we use only the pressure component to guide the aggregation
and neglect all other unknowns on the fine level. All other level are formed
in the usual way by scalar aggregation.
Currently,it has to be requested for flow_ebos manually by passing
"linear_solver_use_amg=true amg_blackoil_system=true" to it.
The wells, FIP and initial output of NNCs is still handled
by code in opm-simulators. The plan is to move more of the
functionality to ebos.
All tests pass and MPI restart works
it seems like most build systems pass a -DHAVE_CONFIG_H flag to the
compiler which still causes `#if HAVE_CONFIG_H` to be false while it
clearly is supposed to be triggered.
That said, I do not really see a good reason why the inclusion of the
`config.h` file should be guarded in the first place: the file is
guaranteed to always available by proper build systems, and if it was
not included the build either breaks at the linking stage or -- at the
very least -- the runtime behavior of the resulting libraries will be
very awkward.
if PETSc is not available, the .cpp file will compile fine because it
will be reduced to be empty, but trying to include
LinearSolverPetsc.hpp in this case will result in an error.
This is quite a hack: Even though energy is not a "phase" and it is
also not considered in MaxNumPhases and pu.num_phases because this
would break a lot of assumptions in old code, it is nevertheless
assigned an "canonical index" that can be translated "active index"
via PhaseUsage::phase_pos[]. This awkwardness is needed because much
of the legacy OPM code conflates the concepts of "fluid phase" and
"conserved quantity" and fixing that issue would basically mean an
almost complete rewrite of much of the legacy code. That said, the
same statement applies to polymer and solvent, but these are currently
handled as even more second-class citizens because they are not even
given a canonical index and also cannot be translated into an active
one.
