Previously, if the problem occured on an MPI process with rank other
than zero the the logging would not seen (at least in the output
files). Now together with the previous commit the problem should be
logged together with the well name and calling method.
For multi segment well the underlying call to
MultisegmentWell::updateWellStateWithTarget (at least if
updateWellStateWithTHPTargetProd is called for a producer under thp
control) might throw as there might be a singular matrix during the
solve needed in MultisegmentWell::iterateWellEqWithControl.
Previously, if that happened then the MPI process where it happened would
stop the nonlinear iteration as failed and try with a chopped time
step. The others might go one with the current time step and we would
see MPI errors about truncated messages.
Now we communicate any exception happening during this part of
WellModel::updateAndCommunicate and all processes will stop the
nonlinear iteration as failed and chop the time step.
We are experiencing singular matrices when solving mulisegment wells
sometimes. In that case (here during
BlackoilModelEbos::assembleReservoir <-
BlackoilModelEbos::initialLinerization <-
BlackoilModelEbos:::nonlinearIterationNewton ) an exception is thrown
when updating the controls of a well.
The problem here is that this exception only happens on one
process. That one goes to the catch block in
NonLinearSolverEbos::step, marks the nonlinear solve as failed and
cuts the time step. The others move to the collective communication
below. Somehow and somewhen all end up in a non-matching collective
communication with different data types and we get an MPI Error that
the message was truncated.
Now all processes will throw, terminate the nonlinear solver and cut
the timestep as it should be.
at the early stage of computeWellRatesWithBhpIterations. The perforation rates are not updated,
and it is not sensible to update based on the inconsistent well rates and perforation rates.
Better to keep the original explicit quantities for better consistency.
Furthermore, it can be dangerous to update the explicit quantities based on the
irrelevant perforation rates, since the ratios can be very undesirable due to crossflow.
Using bool here is at least frowned upon. To be honest, I have no idea
what happens underneath here if we pass a bool. In contrast to other
pod types we do not associate it with a builtin type of MPI (not even
sure what to use). Hence we probably create a custom type for sending
and receiving. That should work. But I have no idea what will be used
for summation.
BTW: I am debugging a case that previously crashed and now suddenly
works and this seems to be the only relevant change I made in the
meantime.
Implement graphcoloring to expose rows in level sets that that can be
executed in parallel during the sparse triangular solves.
Add copy of A matrix that is reordered to ensure continuous memory reads
when traversing the matrix in level set order.
TODO: add number of threads available as constructor argument in DILU
This commit adds support for loading a three-column NLDD
partitioning scheme of the form
MPI_Rank Cartesian_Index NLDD_Domain_ID
from a text file. In an MPI run it is assumed that the first column
holds integers in the range 0..MPI_Size()-1, and typically that each
such integer is listed at least once. In a sequential run, the MPI
rank column is ignored.
With this scheme we can load the same partition files that we write,
for increased repeatability and determinism, and we can also
experiment with externally generated NLDD partitions.
This commit adds a new (hidden) debugging option,
DebugEmitCellPartition (--debug-emit-cell-partition)
which, when set, will cause each rank to write a three-column text
file of the form
MPI_Rank Cartesian_Index NLDD_Domain_ID
into the directory
partition/CaseName
of the run's output directory. That file will be named according to
the process' MPI rank, so the first column will be the same as the
file name.
The option is primarily intended for debugging the NLDD partitioning
scheme, so is mostly reserved for runs with low MPI sizes (e.g.,
less than 20).
While here, also make the MPIPartitionFromFile helper class aware of
this format so that we can use concatenated output files as an input
to the MPI partitioning algorithm for repeatability.
This commit introduces new, experimental support for loading a
partitioning of the cells from a text file. The name of the file is
passed into the simulator using the new, hidden, command line option
--external-partition=filename
and we perform some basic checking that the number of elements in the
partition matches the number of cells in the CpGrid object.
The well dfactor is scaled by the well index
If postive the connection dfactor is threated as a well factor
and also scaled. If negative the connection dfactor is not scaled
This commit switches the current implementation of
'partitionCellsZoltan()', i.e., 'partitionCells("zoltan", ...)' into
using the MPI-aware ParallelNLDDPartitioningZoltan utility. In
doing so we make 'partitionCellsZoltan()' private since its
availability is not guaranteed. We also slightly reorder the
parameters and switch from passing a "Grid" into passing a
"GridView" as an argument to partitionCells(), and specialise this
function for the known grid views in OPM Flow.
We extract the Zoltan-related parameters out to an Entity-dependent
helper structure and move the complexity of forming this type to a
new helper function, BlackoilModelEbosNldd::partitionCells().
Invokes Zoltan library and requires MPI. Client code constructs an
abstract connectivity graph by defining connections/edges through
the 'registerConnection()' member function. May also impose a
restriction that certain cells/vertices be placed in the same
domain/block in the resulting partition. Client code must supply a
callback function that defines globally unique cell/vertex/object
IDs, across all MPI ranks, for each vertex in the connectivity
graph.
Member function 'partitionElement()' forms the resulting partition
vector, the size of which is the total number of objects visible to
the local rank-typically the number of cells owned by the rank, and
the number of overlap cells--i.e., the size of the local grid view.
