inline double Score::calculate_determinant(std::vector <std::vector<double> > M, int n){
Decomposition D;
double determinant = 1.0;
determinant = D.determinant(M, n);
return determinant;
}
void evalBestFitnessPlanDump::call( Decomposition & eo )
{
if( eo.fitness() > _best ) {
_best = eo.fitness();
dump( eo );
}
}
void evalBestMakespanPlanDump::call( Decomposition & eo )
{
if( eo.plan().makespan() < _best ) {
_best = eo.plan().makespan();
dump( eo );
}
}
TEST_F(DecompositionTest, IsNotRootByDefault)
{
Decomposition d = {DecompositionNode{{}}, solverFactory};
EXPECT_FALSE(d.isRoot());
d.setRoot();
EXPECT_TRUE(d.isRoot());
}
void Balancer::perfBalanceGPU(Cluster &cluster, Decomposition& decomp,
const double kTimeEstimate) {
const int kGPUOnly = 2;
WorkQueue work_queue;
WorkRequest work_request;
const int kNumTotalGPUs = cluster.getNumTotalGPUs();
if (decomp.getNumSubDomains() == 0 || kNumTotalGPUs == 0)
return;
for (int gpu_index = 0; gpu_index < kNumTotalGPUs; ++gpu_index) {
Node& gpu = cluster.getGlobalGPU(gpu_index);
// fastest gpu will have largest weight, and thus move to front of queue
work_request.setTimeDiff(kTimeEstimate - gpu.getBalTimeEst(1, kGPUOnly));
work_request.setIndex(gpu_index);
work_queue.push(work_request);
}
const int kNumBlocks = decomp.getNumSubDomains();
// place data blocks on gpu's one-at-a-time
for (int block_id = 0; block_id < kNumBlocks; ++block_id) {
work_request = work_queue.top();
work_queue.pop();
Node& gpu = cluster.getGlobalGPU(work_request.getIndex());
gpu.incrementBalCount();
double time_diff = gpu.getBalTimeEst(1, kGPUOnly);
work_request.setTimeDiff(time_diff);
work_queue.push(work_request);
//printWorkQueue(work_queue);
}
cluster.distributeBlocks(&decomp);
}
RandomAccessIterator2 reduce_intervals(RandomAccessIterator1 first, Decomposition decomp, RandomAccessIterator2 result, BinaryFunction binary_op)
{
typedef typename thrust::iterator_value<RandomAccessIterator2>::type result_type;
const size_t groupsize = 128;
size_t heap_size = groupsize * sizeof(result_type);
bulk::async(bulk::grid<groupsize,7>(decomp.size(),heap_size), reduce_intervals_kernel(), bulk::root.this_exec, first, decomp, result, binary_op);
return result + decomp.size();
} // end reduce_intervals()
TEST_F(DecompositionTest, SetsParent)
{
Decomposition d = {DecompositionNode{{}}, solverFactory};
EXPECT_EQ(nullptr, d.getParent());
DecompositionPtr c{new Decomposition{DecompositionNode{{}}, solverFactory}};
c->setParent(&d);
d.addChild(std::move(c));
EXPECT_EQ(&d, (*d.getChildren().begin())->getParent());
}
__host__ __device__
RandomAccessIterator2 reduce_intervals_(execution_policy<DerivedPolicy> &exec, RandomAccessIterator1 first, Decomposition decomp, RandomAccessIterator2 result, BinaryFunction binary_op)
{
typedef typename thrust::iterator_value<RandomAccessIterator2>::type result_type;
const size_t groupsize = 128;
size_t heap_size = groupsize * sizeof(result_type);
bulk_::async(bulk_::grid<groupsize,7>(decomp.size(),heap_size,stream(thrust::detail::derived_cast(exec))), reduce_intervals_detail::reduce_intervals_kernel(), bulk_::root.this_exec, first, decomp, result, binary_op);
return result + decomp.size();
} // end reduce_intervals()
RandomAccessIterator2 reduce_intervals_(execution_policy<DerivedPolicy> &, RandomAccessIterator1 first, Decomposition decomp, RandomAccessIterator2 result, BinaryFunction binary_op)
{
namespace bulk_ = thrust::system::cuda::detail::bulk;
typedef typename thrust::iterator_value<RandomAccessIterator2>::type result_type;
const size_t groupsize = 128;
size_t heap_size = groupsize * sizeof(result_type);
bulk_::async(bulk_::grid<groupsize,7>(decomp.size(),heap_size), reduce_intervals_detail::reduce_intervals_kernel(), bulk_::root.this_exec, first, decomp, result, binary_op);
return result + decomp.size();
} // end reduce_intervals()
void declareDecomposition(const Decomposition& decomposition, std::ostream& out)
{
out << "% Decomposition facts" << std::endl;
out << "currentNode(" << decomposition.getNode().getGlobalId() << ")." << std::endl;
for(const auto& v : decomposition.getNode().getBag()) {
out << "bag(" << decomposition.getNode().getGlobalId() << ',' << v << "). ";
out << "current(" << v << ")." << std::endl;
}
out << "#const numChildNodes=" << decomposition.getChildren().size() << '.' << std::endl;
if(decomposition.getChildren().empty())
out << "initial." << std::endl;
else {
for(const auto& child : decomposition.getChildren()) {
out << "childNode(" << child->getNode().getGlobalId() << ")." << std::endl;
for(const auto& v : child->getNode().getBag()) {
out << "bag(" << child->getNode().getGlobalId() << ',' << v << "). ";
out << "-introduced(" << v << ")." << std::endl; // Redundant
}
}
}
if(decomposition.isRoot())
out << "final." << std::endl;
if(decomposition.isPostJoinNode())
out << "postJoin." << std::endl;
// Redundant predicates for convenience...
out << "introduced(X) :- current(X), not -introduced(X)." << std::endl;
out << "removed(X) :- childNode(N), bag(N,X), not current(X)." << std::endl;
}
Solver::Solver(const Decomposition& decomposition, const Application& app, const std::vector<std::string>& encodingFiles, bool tableMode, bool cardinalityCost, bool printStatistics)
: ::Solver(decomposition, app)
, encodingFiles(encodingFiles)
, tableMode(tableMode)
, cardinalityCost(cardinalityCost)
, printStatistics(printStatistics)
{
Gringo::message_printer()->disable(Gringo::W_ATOM_UNDEFINED);
#ifndef DISABLE_CHECKS
// TODO: Implement tables::EncodingChecker
if(!tableMode) {
// Check the encoding, but only in the decomposition root.
// Otherwise we'd probably do checks redundantly.
if(decomposition.isRoot()) {
std::ofstream dummyStream;
std::unique_ptr<Gringo::Output::OutputBase> out(new Gringo::Output::OutputBase({}, dummyStream));
Gringo::Input::Program program;
asp_utils::DummyGringoModule module;
Gringo::Scripts scripts(module);
Gringo::Defines defs;
std::unique_ptr<EncodingChecker> encodingChecker{new trees::EncodingChecker(scripts, program, *out, defs)};
Gringo::Input::NonGroundParser parser(*encodingChecker);
for(const auto& file : encodingFiles)
parser.pushFile(std::string(file));
parser.parse();
encodingChecker->check();
}
}
#endif
}
void DebugMachineReadable::solverInvocationResult(const Decomposition& decompositionNode, const ItemTree* result)
{
const auto id = decompositionNode.getNode().getGlobalId();
if(result) {
std::cout << "% Facts describing the resulting item tree at node " << id << std::endl;
std::ostringstream rootItemSetName;
rootItemSetName << 'n' << id;
solver::asp::Solver::declareItemTree(std::cout, result, false, id, rootItemSetName.str());
std::cout << std::endl;
std::cout << "% Memory locations of the item tree nodes at decomposition node " << id << " (not passed to ASP)" << std::endl;
declareItemTreeNodeMemoryAddresses(std::cout, result, rootItemSetName.str());
std::cout << std::endl;
std::cout << "% Extension pointers at decomposition node " << id << " (not passed to ASP)" << std::endl;
declareExtensionPointers(std::cout, result, rootItemSetName.str());
std::cout << std::endl;
std::cout << "% (Derived) costs of non-leaf nodes of the item tree at decomposition node " << id << " (not passed to ASP)" << std::endl;
declareDerivedCosts(std::cout, result, rootItemSetName.str());
std::cout << std::endl;
}
else
std::cout << "% Item tree of node " << id << " is empty." << std::endl;
}
std::unique_ptr<Solver> AspFactory::newSolver(const Decomposition& decomposition) const
{
if(optDefaultJoin.isUsed() && decomposition.isJoinNode())
return std::unique_ptr<Solver>(new DefaultJoin(decomposition, app));
else
return std::unique_ptr<Solver>(new Asp(decomposition, app, optEncodingFile.getValue(), optTables.isUsed()));
}
bool MutationDelGoal::operator()( Decomposition & decompo )
{
if( decompo.size() <= 1 ) {
return false;
} else {
#ifndef NDEBUG
eo::log << eo::debug << "D";
eo::log.flush();
eo::log << eo::xdebug << " DelGoal:" << std::endl << "\tBefore: ";
simplePrint( eo::log << eo::xdebug, decompo );
#endif
unsigned int i = rng.random( std::min( static_cast<unsigned int>(decompo.size()), static_cast<unsigned int>(decompo.last_reached() + 1) ) );
decompo.erase( decompo.iter_at( i ) );
#ifndef NDEBUG
eo::log << eo::xdebug << "\tdelete the " << i << "th goal" << std::endl;
eo::log << eo::xdebug << "\tAfter: ";
simplePrint( eo::log << eo::xdebug, decompo );
#endif
return true;
}
}
void Balancer::balance(Cluster &cluster, Decomposition& decomp,
const int kConfig) {
const int kCPUAndGPU = 0;
const int kCPUOnly = 1;
const int kGPUOnly = 2;
int blocks_per_node = 0;
// num cpu nodes
unsigned int total_nodes = cluster.getNumNodes();
size_t num_blocks = decomp.getNumSubDomains();
// initialize block directory
cluster.setNumBlocks(num_blocks);
if (kConfig != kCPUOnly) {
unsigned int num_gpus = 0;
for (unsigned int node_index = 0;
node_index < cluster.getNumNodes();
++node_index) {
num_gpus += cluster.getNode(node_index).getNumChildren();
}
if (kConfig == kGPUOnly) // gpu only
total_nodes = num_gpus;
else if (kConfig == kCPUAndGPU) // cpu and gpu
total_nodes += num_gpus;
}
blocks_per_node = ceil(decomp.getNumSubDomains() / (float) total_nodes);
for (unsigned int node_index = 0;
node_index < cluster.getNumNodes();
++node_index) {
Node& node = cluster.getNode(node_index);
if (kConfig == kCPUOnly || kConfig == kCPUAndGPU) {
for (int subd = 0;
subd < blocks_per_node && 0 < decomp.getNumSubDomains();
++subd) {
SubDomain *block = decomp.popSubDomain();
node.addSubDomain(block);
}
}
if (kConfig == kGPUOnly || kConfig == kCPUAndGPU) {
for (unsigned int gpu_index = 0;
gpu_index < node.getNumChildren();
++gpu_index) {
Node& gpu = node.getChild(gpu_index);
for (int subd = 0;
subd < blocks_per_node && 0 < decomp.getNumSubDomains();
++subd) {
SubDomain *block = decomp.popSubDomain();
gpu.addSubDomain(block);
}
}
}
}
/* the work is balanced, so we can fill the block directory */
cluster.storeBlockLocs();
}
void Solver::startSolvingForCurrentRowCombination()
{
// ++solverSetups;
asyncResult.reset();
if(reground) {
// Set up ASP solver
config.solve.numModels = 0;
// TODO The last parameter of clasp.startAsp in the next line is "allowUpdate". Does setting it to false have benefits?
// WORKAROUND for BUG in ClaspFacade::startAsp()
// TODO remove on update to new version
if(clasp.ctx.numVars() == 0 && clasp.ctx.frozen())
clasp.ctx.reset();
Clasp::Asp::LogicProgram& claspProgramBuilder = static_cast<Clasp::Asp::LogicProgram&>(clasp.startAsp(config));
GringoOutputProcessor gringoOutput(claspProgramBuilder);
std::unique_ptr<Gringo::Output::OutputBase> out(new Gringo::Output::OutputBase({}, gringoOutput));
Gringo::Input::Program program;
asp_utils::DummyGringoModule module;
Gringo::Scripts scripts(module);
Gringo::Defines defs;
Gringo::Input::NongroundProgramBuilder gringoProgramBuilder(scripts, program, *out, defs);
Gringo::Input::NonGroundParser parser(gringoProgramBuilder);
// Input: Induced subinstance
std::unique_ptr<std::stringstream> instanceInput(new std::stringstream);
asp_utils::induceSubinstance(*instanceInput, app.getInstance(), decomposition.getNode().getBag());
app.getPrinter().solverInvocationInput(decomposition, instanceInput->str());
// Input: Decomposition
std::unique_ptr<std::stringstream> decompositionInput(new std::stringstream);
asp_utils::declareDecomposition(decomposition, *decompositionInput);
app.getPrinter().solverInvocationInput(decomposition, decompositionInput->str());
// Input: Child rows
std::unique_ptr<std::stringstream> childRowsInput(new std::stringstream);
*childRowsInput << "% Child row facts" << std::endl;
for(const auto& row : getCurrentRowCombination()) {
for(const auto& item : row->getItems())
*childRowsInput << "childItem(" << item << ")." << std::endl;
for(const auto& item : row->getAuxItems())
*childRowsInput << "childAuxItem(" << item << ")." << std::endl;
// TODO costs, etc.
}
app.getPrinter().solverInvocationInput(decomposition, childRowsInput->str());
// Pass input to ASP solver
for(const auto& file : encodingFiles)
parser.pushFile(std::string(file));
parser.pushStream("<instance>", std::move(instanceInput));
parser.pushStream("<decomposition>", std::move(decompositionInput));
parser.pushStream("<child_rows>", std::move(childRowsInput));
parser.parse();
// Ground
program.rewrite(defs);
program.check();
if(Gringo::message_printer()->hasError())
throw std::runtime_error("Grounding stopped because of errors");
auto gPrg = program.toGround(out->domains);
Gringo::Ground::Parameters params;
params.add("base", {});
gPrg.ground(params, scripts, *out);
params.clear();
claspProgramBuilder.endProgram();
itemAtomInfos.clear();
for(const auto& atom : gringoOutput.getItemAtomInfos())
itemAtomInfos.emplace_back(ItemAtomInfo(atom, claspProgramBuilder));
auxItemAtomInfos.clear();
for(const auto& atom : gringoOutput.getAuxItemAtomInfos())
auxItemAtomInfos.emplace_back(AuxItemAtomInfo(atom, claspProgramBuilder));
// TODO costs etc.
clasp.prepare();
}
else {
// Set external variables to the values of the current child row combination
clasp.update(false, false);
clasp.prepare();
// Mark atoms corresponding to items from the currently extended rows
for(const auto& row : getCurrentRowCombination()) {
for(const auto& item : row->getItems()) {
assert(itemsToLitIndices.find(item) != itemsToLitIndices.end());
assert(itemsToLitIndices.at(item) < literals.size());
#ifdef DISABLE_CHECKS
literals[itemsToLitIndices.at(item)].watch();
#else
try {
literals[itemsToLitIndices.at(item)].watch();
}
catch(const std::out_of_range&) {
std::ostringstream msg;
msg << "Unknown variable; atom childItem(" << *item << ") not shown or not declared as external?";
throw std::runtime_error(msg.str());
}
#endif
}
for(const auto& item : row->getAuxItems()) {
assert(auxItemsToLitIndices.find(item) != auxItemsToLitIndices.end());
assert(auxItemsToLitIndices.at(item) < literals.size());
#ifdef DISABLE_CHECKS
literals[auxItemsToLitIndices.at(item)].watch();
#else
try {
literals[auxItemsToLitIndices.at(item)].watch();
}
catch(const std::out_of_range&) {
std::ostringstream msg;
msg << "Unknown variable; atom childAuxItem(" << *item << ") not shown or not declared as external?";
throw std::runtime_error(msg.str());
}
#endif
}
}
// Set marked literals to true and all others to false
for(auto& lit : literals) {
if(lit.watched()) {
lit.clearWatch();
clasp.assume(lit);
}
else
clasp.assume(~lit);
}
}
asyncResult.reset(new BasicSolveIter(clasp));
}
void Balancer::perfBalance(Cluster &cluster, Decomposition& decomp,
const int kConfig) {
WorkQueue work_queue;
WorkRequest work_request;
double total_weight(0.0);
double min_edge_weight(0.0);
double procTime(0.0);
double commTime(0.0);
double timeEst = procTime;
bool changed(false);
const int kGPUOnly(2);
const int kStrongest(3);
Node &root = cluster.getNode(0);
size_t num_blocks = decomp.getNumSubDomains();
// initialize block directory
cluster.setNumBlocks(num_blocks);
//get total iterations per second for cluster
for (unsigned int node = 0; node < cluster.getNumNodes(); ++node) {
total_weight += cluster.getNode(node).getTotalWeight(kConfig);
min_edge_weight += cluster.getNode(node).getMinEdgeWeight(kConfig);
}
// quick estimation of runtime
procTime = num_blocks / total_weight;
commTime = num_blocks / min_edge_weight;
timeEst = procTime;
if (0.0 < min_edge_weight)
timeEst += commTime;
if (kGPUOnly == kConfig) {
perfBalanceGPU(cluster, decomp, timeEst);
} else if (kStrongest == kConfig) {
perfBalanceStrongestDevice(cluster, decomp);
} else {
// perform initial task distribution
for (size_t i = 0; i < num_blocks; ++i)
root.incrementBalCount();
/*fprintf(stderr,
"perfBalance: \n\ttime est: %f sec\n\tprocTime: %f sec\n\tcommTime: %f \
sec\n\ttotal weight:%e \n\tmin edge weight:%e.\n",
timeEst, procTime, commTime, total_weight, min_edge_weight); // */
do {
changed = false;
// balance the work between nodes and root
for (unsigned int cpu_index = 1;
cpu_index < cluster.getNumNodes();
++cpu_index) {
Node& cpu_node = cluster.getNode(cpu_index);
int work_deficit = cpu_node.getTotalWorkNeeded(timeEst, kConfig) -
cpu_node.getBalCount();
if (0 > work_deficit) { // node has extra work
int extra_blocks = abs(work_deficit);
for (int block_index = 0;
(block_index < extra_blocks) &&
(0 < cpu_node.getBalCount());
++block_index) {
// move block from child to parent
cpu_node.decrementBalCount();
root.incrementBalCount();
changed = true;
}
} else if (0 < work_deficit) { //child needs more work
work_request.setTimeDiff(timeEst - cpu_node.getBalTimeEst(0, kConfig));
work_request.setIndex(cpu_index);
work_queue.push(work_request);
}
}
for (unsigned int cpu_index = 0;
cpu_index < root.getNumChildren();
++cpu_index) {
Node& cpu_node = root.getChild(cpu_index);
int work_deficit = cpu_node.getTotalWorkNeeded(timeEst, kConfig) -
cpu_node.getBalCount();
if (0 > work_deficit) { // child has extra work
int extra_blocks = abs(work_deficit);
for (int block_index = 0;
(block_index < extra_blocks) && (0 < cpu_node.getBalCount());
++block_index) {
// move block from child to parent
cpu_node.decrementBalCount();
root.incrementBalCount();
changed = true;
}
} else if (0 < work_deficit) { // child needs more work
work_request.setTimeDiff(timeEst - cpu_node.getBalTimeEst(0, kConfig));
work_request.setIndex(-1 * cpu_index); // hack so I know to give to one of root's children
work_queue.push(work_request);
}
}
/*
at this point we have all extra blocks, and
now we need to distribute blocks to children
that need it
*/
while (0 < root.getBalCount() && // there are blocks left to give
!work_queue.empty()) { // there are requests left to fill
// get largest request
WorkRequest tmp = work_queue.top();
work_queue.pop();
double newTimeDiff = 0.0;
int id = tmp.getIndex();
if (id <= 0) { // local child
id = -1 * id;
root.decrementBalCount();
root.getChild(id).incrementBalCount();
newTimeDiff = timeEst - root.getChild(id).getBalTimeEst(0, kConfig);
changed = true;
} else { // request was from another node in cluster
root.decrementBalCount();
cluster.getNode(id).incrementBalCount();
newTimeDiff = timeEst - cluster.getNode(id).getBalTimeEst(0, kConfig);
changed = true;
}
// if there is still work left to do put it back on
// the queue so that it will reorder correctly
if (0 < newTimeDiff) {
tmp.setTimeDiff(newTimeDiff);
work_queue.push(tmp);
}
}
// balance the work within each node
for (unsigned int node = 0; node < cluster.getNumNodes(); ++node) {
changed |= balanceNode(cluster.getNode(node), timeEst, kConfig);
}
} while (changed);
}
/* now that we know where everything should go, distribute the blocks */
cluster.distributeBlocks(&decomp);
/* the work is balanced, so we can fill the block directory */
cluster.storeBlockLocs();
}
void DebugMachineReadable::solverInvocationInput(const Decomposition& decompositionNode, const std::string& input)
{
std::cout << "% Input for solver at decomposition node " << decompositionNode.getNode().getGlobalId() << std::endl << input << std::endl;
}
Solver::Solver(const Decomposition& decomposition, const Application& app, const std::vector<std::string>& encodingFiles, bool reground, BranchAndBoundLevel bbLevel)
: ::LazySolver(decomposition, app, bbLevel)
, reground(reground)
, encodingFiles(encodingFiles)
{
Gringo::message_printer()->disable(Gringo::W_ATOM_UNDEFINED);
if(!reground) {
// Set up ASP solver
config.solve.numModels = 0;
Clasp::Asp::LogicProgram& claspProgramBuilder = static_cast<Clasp::Asp::LogicProgram&>(clasp.startAsp(config, true)); // TODO In leaves updates might not be necessary.
struct LazyGringoOutputProcessor : GringoOutputProcessor
{
LazyGringoOutputProcessor(Solver* s, Clasp::Asp::LogicProgram& prg)
: GringoOutputProcessor(prg), self(s)
{
}
void storeAtom(unsigned int atomUid, Gringo::Value v) override
{
const std::string& n = *v.name();
if(n == "childItem") {
ASP_CHECK(v.args().size() == 1, "'childItem' predicate does not have arity 1");
std::ostringstream argument;
v.args().front().print(argument);
self->itemsToLitIndices.emplace(String(argument.str()), self->literals.size());
self->literals.push_back(Clasp::posLit(atomUid));
}
else if(n == "childAuxItem") {
ASP_CHECK(v.args().size() == 1, "'childAuxItem' predicate does not have arity 1");
std::ostringstream argument;
v.args().front().print(argument);
self->auxItemsToLitIndices.emplace(String(argument.str()), self->literals.size());
self->literals.push_back(Clasp::posLit(atomUid));
}
GringoOutputProcessor::storeAtom(atomUid, v);
}
Solver* self;
} gringoOutput(this, claspProgramBuilder);
std::unique_ptr<Gringo::Output::OutputBase> out(new Gringo::Output::OutputBase({}, gringoOutput));
Gringo::Input::Program program;
asp_utils::DummyGringoModule module;
Gringo::Scripts scripts(module);
Gringo::Defines defs;
Gringo::Input::NongroundProgramBuilder gringoProgramBuilder(scripts, program, *out, defs);
Gringo::Input::NonGroundParser parser(gringoProgramBuilder);
// Input: Induced subinstance
std::unique_ptr<std::stringstream> instanceInput(new std::stringstream);
asp_utils::induceSubinstance(*instanceInput, app.getInstance(), decomposition.getNode().getBag());
app.getPrinter().solverInvocationInput(decomposition, instanceInput->str());
// Input: Decomposition
std::unique_ptr<std::stringstream> decompositionInput(new std::stringstream);
asp_utils::declareDecomposition(decomposition, *decompositionInput);
app.getPrinter().solverInvocationInput(decomposition, decompositionInput->str());
// Pass input to ASP solver
for(const auto& file : encodingFiles)
parser.pushFile(std::string(file));
parser.pushStream("<instance>", std::move(instanceInput));
parser.pushStream("<decomposition>", std::move(decompositionInput));
parser.parse();
// Ground
program.rewrite(defs);
program.check();
if(Gringo::message_printer()->hasError())
throw std::runtime_error("Grounding stopped because of errors");
auto gPrg = program.toGround(out->domains);
Gringo::Ground::Parameters params;
params.add("base", {});
gPrg.ground(params, scripts, *out);
params.clear();
// Set value of external atoms to free
for(const auto& p : literals)
claspProgramBuilder.freeze(p.var(), Clasp::value_free);
// Finalize ground program and create solver literals
claspProgramBuilder.endProgram();
// Map externals to their solver literals
for(auto& p : literals) {
p = claspProgramBuilder.getLiteral(p.var());
assert(!p.watched()); // Literal must not be watched
}
for(const auto& atom : gringoOutput.getItemAtomInfos())
itemAtomInfos.emplace_back(ItemAtomInfo(atom, claspProgramBuilder));
for(const auto& atom : gringoOutput.getAuxItemAtomInfos())
auxItemAtomInfos.emplace_back(AuxItemAtomInfo(atom, claspProgramBuilder));
// for(const auto& atom : gringoOutput->getCurrentCostAtomInfos())
// currentCostAtomInfos.emplace_back(CurrentCostAtomInfo(atom, claspProgramBuilder));
// for(const auto& atom : gringoOutput->getCostAtomInfos())
// costAtomInfos.emplace_back(CostAtomInfo(atom, claspProgramBuilder)));
// Prepare for solving.
clasp.prepare();
}
}
LeafSolver::LeafSolver(const Decomposition& decomposition, const Application& app)
: ::Solver(decomposition, app)
{
if(decomposition.getNode().getBag().empty() == false)
throw std::runtime_error("ASP solver requires empty leaves");
}
void Balancer::perfBalanceStrongestDevice(Cluster &cluster, Decomposition& decomp) {
WorkQueue work_queue;
WorkRequest work_request;
double total_weight(0.0);
double min_edge_weight(0.0);
double procTime(0.0);
double commTime(0.0);
double timeEst = procTime;
bool changed(false);
const int kStrongest(3);
const int kConfig = kStrongest;
Node &root = cluster.getNode(0);
const size_t kNumBlocks = decomp.getNumSubDomains();
// initialize block directory
cluster.setNumBlocks(kNumBlocks);
//get total iterations per second for cluster
for (unsigned int node = 0; node < cluster.getNumNodes(); ++node) {
total_weight += cluster.getNode(node).getTotalWeight(kConfig);
min_edge_weight += cluster.getNode(node).getMinEdgeWeight(kConfig);
}
// quick estimation of runtime
procTime = (0.0 == total_weight) ?
std::numeric_limits<double>::max() :
kNumBlocks / total_weight;
commTime = (0.0 == min_edge_weight) ?
std::numeric_limits<double>::max() :
kNumBlocks / min_edge_weight;
timeEst = procTime;
timeEst += (std::numeric_limits<double>::max() - procTime >= commTime) ?
commTime :
0.0;
//printf("timeEst:%f\n", timeEst);
// place all of the blocks on the root node
for (size_t i = 0; i < kNumBlocks; ++i)
root.incrementBalCount();
/*fprintf(stderr,
"perfBalance: \n\ttime est: %f sec\n\tprocTime: %f sec\n\tcommTime: %f \
sec\n\ttotal weight:%e \n\tmin edge weight:%e.\n",
timeEst, procTime, commTime, total_weight, min_edge_weight); // */
do {
changed = false;
//printf("beginning, changed == %s\n", (changed == true) ? "true" : "false");
// balance the work between nodes and root
for (unsigned int cpu_index = 1;
cpu_index < cluster.getNumNodes();
++cpu_index) {
Node& cpu_node = cluster.getNode(cpu_index);
int work_deficit = cpu_node.getTotalWorkNeeded(timeEst, kConfig) -
cpu_node.getBalCount();
if (0 > work_deficit) { // node has extra work
//printf("cpu node %d has %d extra blocks.\n", cpu_index, work_deficit);
int extra_blocks = abs(work_deficit);
for (int block_index = 0;
(block_index < extra_blocks) &&
(0 < cpu_node.getBalCount());
++block_index) {
// move block from child to parent
cpu_node.decrementBalCount();
root.incrementBalCount();
changed = true;
}
} else if (0 < work_deficit) { //child needs more work
work_request.setTimeDiff(timeEst - cpu_node.getBalTimeEst(0, kConfig));
work_request.setIndex(cpu_index);
work_queue.push(work_request);
}
}
// go through all of root nodes gpus
for (unsigned int index = 0; index < root.getNumChildren(); ++index) {
Node& node = root.getChild(index);
int work_deficit = node.getTotalWorkNeeded(timeEst, kConfig) -
node.getBalCount();
if (0 > work_deficit) { // child has extra work
//printf("root child %d has %d blocks and only needs %d blocks.\n", index,
// node.getBalCount(), node.getTotalWorkNeeded(timeEst, kConfig));
int extra_blocks = abs(work_deficit);
for (int block_index = 0;
(block_index < extra_blocks) && (0 < node.getBalCount());
++block_index) {
// move block from child to parent
node.decrementBalCount();
root.incrementBalCount();
//changed = true;
}
} else if (0 < work_deficit) { // child needs more work
work_request.setTimeDiff(timeEst - node.getBalTimeEst(0, kConfig));
work_request.setIndex(-1 * index); // hack so I know to give to one of root's children
work_queue.push(work_request);
}
}
/*
at this point we have all extra blocks, and
now we need to distribute blocks to children
that need it
*/
//printf("after collecting extra blocks from children, changed == %s\n",
// (changed == true) ? "true" : "false");
// while root has extra blocks and there are requests left to fill
while (0 < (root.getBalCount() - root.getTotalWorkNeeded(timeEst, kConfig)) &&
!work_queue.empty()) {
//printf("root needs %d blocks and has %d blocks.\n",
// root.getTotalWorkNeeded(timeEst, kConfig), root.getBalCount());
// get largest request
WorkRequest tmp = work_queue.top();
work_queue.pop();
double newTimeDiff = 0.0;
int id = tmp.getIndex();
if (id <= 0) { // local child
//printf("giving block to local child.\n");
id = -1 * id;
root.decrementBalCount();
root.getChild(id).incrementBalCount();
newTimeDiff = timeEst - root.getChild(id).getBalTimeEst(0, kConfig);
changed = true;
} else { // request was from another node in cluster
//printf("giving block to cpu node child.\n");
root.decrementBalCount();
cluster.getNode(id).incrementBalCount();
newTimeDiff = timeEst - cluster.getNode(id).getBalTimeEst(0, kConfig);
changed = true;
}
// if there is still work left to do put it back on
// the queue so that it will reorder correctly
if (0 < newTimeDiff) {
tmp.setTimeDiff(newTimeDiff);
work_queue.push(tmp);
}
}
//printf("after distributing extra blocks to cpu nodes, changed == %s\n",
// (changed == true) ? "true" : "false");
// balance the work within each node
for (unsigned int node = 0; node < cluster.getNumNodes(); ++node) {
balanceNode(cluster.getNode(node), timeEst, kConfig);
//changed |= balanceNode(cluster.getNode(node), timeEst, kConfig);
}
//printf("after balancing within each node, changed == %s\n",
// (changed == true) ? "true" : "false");
//printClusterBalCount(cluster);
//printf("************* END OF BALANCE ITERATION ***********\n");
} while (changed);
}
