mirror of
https://github.com/OPM/opm-simulators.git
synced 2025-02-25 18:55:30 -06:00
Fixes a bug in GasLiftStage2_impl.hpp
When making gas lift parallel, see PR #3148, redistributeALQ() did not reserve space for the decremental and incremental gradients. Later in execution thread when push_back() was called to add elements to the vectors, the capacity of the vector could get exceeded and hence the internal representation of the vector could be reallocated. This seems to have caused access to undefined memory errors since the iterators into the vectors was long longer valid.
This commit is contained in:
Håkon Hægland
parent
842e0a53a4
commit
0abcac3777
@ -69,6 +69,12 @@ namespace Opm
|
||||
using GradPairItr = std::vector<GradPair>::iterator;
|
||||
using GradInfo = typename GasLiftSingleWell::GradInfo;
|
||||
using GradMap = std::map<std::string, GradInfo>;
|
||||
using MPIComm = typename Dune::MPIHelper::MPICommunicator;
|
||||
#if DUNE_VERSION_NEWER(DUNE_COMMON, 2, 7)
|
||||
using Communication = Dune::Communication<MPIComm>;
|
||||
#else
|
||||
using Communication = Dune::CollectiveCommunication<MPIComm>;
|
||||
#endif
|
||||
static const int Water = BlackoilPhases::Aqua;
|
||||
static const int Oil = BlackoilPhases::Liquid;
|
||||
static const int Gas = BlackoilPhases::Vapour;
|
||||
@ -127,8 +133,10 @@ namespace Opm
|
||||
const std::string &name, GradInfo &grad, bool increase);
|
||||
void updateGradVector_(
|
||||
const std::string &name, std::vector<GradPair> &grads, double grad);
|
||||
std::vector<GradPair> updateGlobalGradVector_(const std::vector<GradPair> &grads_global) const;
|
||||
std::vector<GradPair> localToGlobalGradVector_(const std::vector<GradPair> &grads_local) const;
|
||||
void mpiSyncGlobalGradVector_(std::vector<GradPair> &grads_global) const;
|
||||
void mpiSyncLocalToGlobalGradVector_(
|
||||
const std::vector<GradPair> &grads_local,
|
||||
std::vector<GradPair> &grads_global) const;
|
||||
|
||||
|
||||
DeferredLogger &deferred_logger_;
|
||||
@ -144,6 +152,7 @@ namespace Opm
|
||||
const Schedule &schedule_;
|
||||
const PhaseUsage &phase_usage_;
|
||||
const GasLiftOpt& glo_;
|
||||
const Communication &comm_;
|
||||
GradMap inc_grads_;
|
||||
GradMap dec_grads_;
|
||||
bool debug_;
|
||||
|
||||
@ -56,6 +56,7 @@ GasLiftStage2(
|
||||
schedule_{ebos_simulator.vanguard().schedule()},
|
||||
phase_usage_{well_model_.phaseUsage()},
|
||||
glo_{schedule_.glo(report_step_idx_)},
|
||||
comm_{ebos_simulator.vanguard().grid().comm()},
|
||||
debug_{false}
|
||||
{
|
||||
// this->time_step_idx_
|
||||
@ -286,8 +287,7 @@ GasLiftStage2<TypeTag>::
|
||||
getCurrentGroupRates_(const Group &group)
|
||||
{
|
||||
auto rates = getCurrentGroupRatesRecursive_(group);
|
||||
const auto& comm = ebos_simulator_.vanguard().grid().comm();
|
||||
comm.sum(rates.data(), rates.size());
|
||||
this->comm_.sum(rates.data(), rates.size());
|
||||
auto [oil_rate, gas_rate, alq] = rates;
|
||||
if (this->debug_) {
|
||||
const std::string msg = fmt::format(
|
||||
@ -468,6 +468,64 @@ getGroupGliftWellsRecursive_(const Group &group,
|
||||
}
|
||||
}
|
||||
|
||||
template<typename TypeTag>
|
||||
void
|
||||
GasLiftStage2<TypeTag>::
|
||||
mpiSyncGlobalGradVector_(std::vector<GradPair> &grads_global) const
|
||||
{
|
||||
if (this->comm_.size() == 1)
|
||||
return;
|
||||
|
||||
std::vector<GradPair> grads_local;
|
||||
for (auto itr = grads_global.begin(); itr != grads_global.end(); itr++) {
|
||||
if (well_state_map_.count(itr->first) > 0) {
|
||||
grads_local.push_back(*itr);
|
||||
}
|
||||
}
|
||||
mpiSyncLocalToGlobalGradVector_(grads_local, grads_global);
|
||||
}
|
||||
|
||||
template<typename TypeTag>
|
||||
void
|
||||
GasLiftStage2<TypeTag>::
|
||||
mpiSyncLocalToGlobalGradVector_(
|
||||
const std::vector<GradPair> &grads_local, std::vector<GradPair> &grads_global) const
|
||||
{
|
||||
assert(this->comm_.size() > 1); // The parent should check if comm. size is > 1
|
||||
using Pair = std::pair<int, double>;
|
||||
std::vector<Pair> grads_local_tmp;
|
||||
grads_local_tmp.reserve(grads_local.size());
|
||||
for (size_t i = 0; i < grads_local.size(); ++i) {
|
||||
if(!this->well_state_.wellIsOwned(grads_local[i].first))
|
||||
continue;
|
||||
grads_local_tmp.push_back(
|
||||
std::make_pair(
|
||||
this->well_state_.wellNameToGlobalIdx(grads_local[i].first),
|
||||
grads_local[i].second));
|
||||
}
|
||||
|
||||
std::vector<int> sizes_(this->comm_.size());
|
||||
std::vector<int> displ_(this->comm_.size() + 1, 0);
|
||||
int mySize = grads_local_tmp.size();
|
||||
this->comm_.allgather(&mySize, 1, sizes_.data());
|
||||
std::partial_sum(sizes_.begin(), sizes_.end(), displ_.begin()+1);
|
||||
std::vector<Pair> grads_global_tmp(displ_.back());
|
||||
|
||||
this->comm_.allgatherv(grads_local_tmp.data(), grads_local_tmp.size(),
|
||||
grads_global_tmp.data(), sizes_.data(), displ_.data());
|
||||
|
||||
// NOTE: This leaves the capacity of 'grads_global' unchanged, so
|
||||
// memory is not reallocated here
|
||||
grads_global.clear();
|
||||
|
||||
for (size_t i = 0; i < grads_global_tmp.size(); ++i) {
|
||||
grads_global.emplace_back(
|
||||
std::make_pair(
|
||||
well_state_.globalIdxToWellName(grads_global_tmp[i].first),
|
||||
grads_global_tmp[i].second));
|
||||
}
|
||||
}
|
||||
|
||||
template<typename TypeTag>
|
||||
void
|
||||
GasLiftStage2<TypeTag>::
|
||||
@ -514,13 +572,16 @@ optimizeGroupsRecursive_(const Group &group)
|
||||
optimizeGroup_(group);
|
||||
|
||||
}
|
||||
|
||||
template<typename TypeTag>
|
||||
void
|
||||
GasLiftStage2<TypeTag>::
|
||||
recalculateGradientAndUpdateData_(
|
||||
GradPairItr &grad_itr, bool increase,
|
||||
std::vector<GradPair> &grads, std::vector<GradPair> &other_grads)
|
||||
|
||||
//incremental and decremental gradients, if 'grads' are incremental, then
|
||||
// 'other_grads' are decremental, or conversely, if 'grads' are decremental, then
|
||||
// 'other_grads' are incremental
|
||||
std::vector<GradPair> &grads, std::vector<GradPair> &other_grads)
|
||||
{
|
||||
// NOTE: We make a copy of the name string instead of taking a reference
|
||||
// since we may have to erase grad_itr (in the "else" condition below)
|
||||
@ -547,6 +608,8 @@ recalculateGradientAndUpdateData_(
|
||||
// The old incremental gradient becomes the new decremental gradient
|
||||
// or the old decremental gradient becomes the new incremental gradient
|
||||
updateGrad_(name, *old_grad, !increase);
|
||||
// NOTE: This may invalidate any iterator into 'other_grads' since
|
||||
// updateGradVector_() will do a push_back() if 'name' is not found..
|
||||
updateGradVector_(name, other_grads, old_grad->grad);
|
||||
}
|
||||
}
|
||||
@ -601,15 +664,24 @@ redistributeALQ_(std::vector<GasLiftSingleWell *> &wells, const Group &group,
|
||||
std::vector<GradPair> &inc_grads, std::vector<GradPair> &dec_grads)
|
||||
{
|
||||
OptimizeState state {*this, group};
|
||||
std::vector<GradPair> inc_grads_local;
|
||||
std::vector<GradPair> dec_grads_local;
|
||||
inc_grads_local.reserve(wells.size());
|
||||
dec_grads_local.reserve(wells.size());
|
||||
state.calculateEcoGradients(wells, inc_grads_local, dec_grads_local);
|
||||
|
||||
// the gradients needs to be communicated to all ranks
|
||||
dec_grads = localToGlobalGradVector_(dec_grads_local);
|
||||
inc_grads = localToGlobalGradVector_(inc_grads_local);
|
||||
// NOTE: 'inc_grads' and 'dec_grads' can never grow larger than wells.size()
|
||||
// By reserving space here, we can ensure that any push_back() on these
|
||||
// will never reallocate memory and invalidate any iterators.
|
||||
inc_grads.reserve(wells.size());
|
||||
dec_grads.reserve(wells.size());
|
||||
if (this->comm_.size() == 1) {
|
||||
state.calculateEcoGradients(wells, inc_grads, dec_grads);
|
||||
}
|
||||
else {
|
||||
std::vector<GradPair> inc_grads_local;
|
||||
std::vector<GradPair> dec_grads_local;
|
||||
inc_grads_local.reserve(wells.size());
|
||||
dec_grads_local.reserve(wells.size());
|
||||
state.calculateEcoGradients(wells, inc_grads_local, dec_grads_local);
|
||||
// the gradients needs to be communicated to all ranks
|
||||
mpiSyncLocalToGlobalGradVector_(dec_grads_local, dec_grads);
|
||||
mpiSyncLocalToGlobalGradVector_(inc_grads_local, inc_grads);
|
||||
}
|
||||
|
||||
if (!state.checkAtLeastTwoWells(wells)) {
|
||||
// NOTE: Even though we here in redistributeALQ_() do not use the
|
||||
@ -705,8 +777,8 @@ removeSurplusALQ_(const Group &group,
|
||||
dec_grad_itr, /*increase=*/false, dec_grads, inc_grads);
|
||||
|
||||
// The dec_grads and inc_grads needs to be syncronized across ranks
|
||||
dec_grads = updateGlobalGradVector_(dec_grads);
|
||||
inc_grads = updateGlobalGradVector_(inc_grads);
|
||||
mpiSyncGlobalGradVector_(dec_grads);
|
||||
mpiSyncGlobalGradVector_(inc_grads);
|
||||
// NOTE: recalculateGradientAndUpdateData_() will remove the current gradient
|
||||
// from dec_grads if it cannot calculate a new decremental gradient.
|
||||
// This will invalidate dec_grad_itr and well_name
|
||||
@ -803,58 +875,6 @@ updateGradVector_(const std::string &name, std::vector<GradPair> &grads, double
|
||||
// later in getEcoGradients()
|
||||
}
|
||||
|
||||
template<typename TypeTag>
|
||||
std::vector<typename GasLiftStage2<TypeTag>::GradPair>
|
||||
GasLiftStage2<TypeTag>::
|
||||
localToGlobalGradVector_(const std::vector<GradPair> &grads_local) const
|
||||
{
|
||||
const auto& comm = ebos_simulator_.vanguard().grid().comm();
|
||||
if (comm.size() == 1)
|
||||
return grads_local;
|
||||
|
||||
using Pair = std::pair<int, double>;
|
||||
std::vector<Pair> grads_local_tmp;
|
||||
grads_local_tmp.reserve(grads_local.size());
|
||||
for (size_t i = 0; i < grads_local.size(); ++i) {
|
||||
if(!well_state_.wellIsOwned(grads_local[i].first))
|
||||
continue;
|
||||
|
||||
grads_local_tmp.push_back(std::make_pair(well_state_.wellNameToGlobalIdx(grads_local[i].first), grads_local[i].second));
|
||||
}
|
||||
|
||||
std::vector<int> sizes_(comm.size());
|
||||
std::vector<int> displ_(comm.size() + 1, 0);
|
||||
int mySize = grads_local_tmp.size();
|
||||
comm.allgather(&mySize, 1, sizes_.data());
|
||||
std::partial_sum(sizes_.begin(), sizes_.end(), displ_.begin()+1);
|
||||
std::vector<Pair> grads_global(displ_.back());
|
||||
|
||||
comm.allgatherv(grads_local_tmp.data(), grads_local_tmp.size(), grads_global.data(), sizes_.data(), displ_.data());
|
||||
std::vector<GradPair> grads(grads_global.size());
|
||||
for (size_t i = 0; i < grads_global.size(); ++i) {
|
||||
grads[i] = std::make_pair(well_state_.globalIdxToWellName(grads_global[i].first), grads_global[i].second);
|
||||
}
|
||||
return grads;
|
||||
}
|
||||
|
||||
template<typename TypeTag>
|
||||
std::vector<typename GasLiftStage2<TypeTag>::GradPair>
|
||||
GasLiftStage2<TypeTag>::
|
||||
updateGlobalGradVector_(const std::vector<GradPair> &grads_global) const
|
||||
{
|
||||
const auto& comm = ebos_simulator_.vanguard().grid().comm();
|
||||
if (comm.size() == 1)
|
||||
return grads_global;
|
||||
|
||||
std::vector<GradPair> grads_local;
|
||||
for (auto itr = grads_global.begin(); itr != grads_global.end(); itr++) {
|
||||
if (well_state_map_.count(itr->first) > 0) {
|
||||
grads_local.push_back(*itr);
|
||||
}
|
||||
}
|
||||
return localToGlobalGradVector_(grads_local);
|
||||
}
|
||||
|
||||
/***********************************************
|
||||
* Public methods declared in OptimizeState
|
||||
***********************************************/
|
||||
@ -894,8 +914,7 @@ checkAtLeastTwoWells(std::vector<GasLiftSingleWell *> &wells)
|
||||
continue;
|
||||
numberOfwells++;
|
||||
}
|
||||
const auto& comm = this->parent.ebos_simulator_.vanguard().grid().comm();
|
||||
numberOfwells = comm.sum(numberOfwells);
|
||||
numberOfwells = this->parent.comm_.sum(numberOfwells);
|
||||
if (numberOfwells < 2) {
|
||||
const std::string msg = fmt::format(
|
||||
"skipping: too few wells ({}) to optimize.", numberOfwells);
|
||||
@ -970,8 +989,8 @@ recalculateGradients(
|
||||
min_dec_grad_itr, /*increase=*/false, dec_grads, inc_grads);
|
||||
|
||||
// The dec_grads and inc_grads needs to be syncronized across ranks
|
||||
dec_grads = this->parent.updateGlobalGradVector_(dec_grads);
|
||||
inc_grads = this->parent.updateGlobalGradVector_(inc_grads);
|
||||
this->parent.mpiSyncGlobalGradVector_(dec_grads);
|
||||
this->parent.mpiSyncGlobalGradVector_(inc_grads);
|
||||
}
|
||||
|
||||
// Take one ALQ increment from well1, and give it to well2
|
||||
@ -1132,8 +1151,7 @@ updateRates(const std::string &well_name)
|
||||
}
|
||||
|
||||
// and communicate the results
|
||||
const auto& comm = this->parent.ebos_simulator_.vanguard().grid().comm();
|
||||
comm.sum(delta.data(), delta.size());
|
||||
this->parent.comm_.sum(delta.data(), delta.size());
|
||||
|
||||
// and update
|
||||
const auto& [delta_oil, delta_gas, delta_alq] = delta;
|
||||
|
||||
Loading…
Reference in New Issue
Block a user