Commit 8783776c authored by Sebastian Eibl's avatar Sebastian Eibl
Browse files

Merge branch 'ParMetisFix' into 'master'

Fix calling of parmetis

See merge request walberla/walberla!88
parents 252cc28b 2c44004b
......@@ -30,20 +30,28 @@
#include "core/mpi/MPIManager.h"
#include "core/mpi/Gather.h"
#include "core/mpi/Gatherv.h"
#include "core/mpi/Reduce.h"
#include "core/timing/Timer.h"
#include <boost/algorithm/string/case_conv.hpp>
#include <boost/algorithm/string/trim.hpp>
#include <array>
#include <vector>
namespace walberla {
namespace blockforest {
std::pair<uint_t, uint_t> getBlockSequenceRange( const PhantomBlockForest & phantomForest, MPI_Comm comm )
std::pair<uint_t, uint_t> getBlockSequenceRange( uint_t numLocalBlocks, MPI_Comm comm )
{
const uint_t rank = uint_c(mpi::translateRank(mpi::MPIManager::instance()->comm(), comm, MPIManager::instance()->rank()));
uint_t numLocalBlocks = phantomForest.getNumberOfBlocks();
WALBERLA_DEBUG_SECTION()
{
int rankRaw;
MPI_Comm_rank(comm, &rankRaw);
WALBERLA_ASSERT_EQUAL(rank, rankRaw);
}
size_t sequenceStartOnProcess = 0;
MPI_Exscan( &numLocalBlocks, &sequenceStartOnProcess, 1, MPITrait<uint_t>::type(), MPI_SUM, comm );
......@@ -53,14 +61,13 @@ std::pair<uint_t, uint_t> getBlockSequenceRange( const PhantomBlockForest & phan
return std::make_pair( sequenceStartOnProcess, sequenceStartOnProcess + numLocalBlocks );
}
std::map< blockforest::BlockID, uint_t > getBlockIdToSequenceMapping( const PhantomBlockForest & phantomForest, const std::pair<uint_t, uint_t> & blockSequenceRange, MPI_Comm comm )
std::map< blockforest::BlockID, uint_t > getBlockIdToSequenceMapping( const PhantomBlockForest& phantomForest, const std::vector< std::pair< const PhantomBlock *, uint_t > > & targetProcess, const std::pair<uint_t, uint_t> & blockSequenceRange, MPI_Comm comm )
{
std::map< blockforest::BlockID, uint_t > mapping;
const auto & blockMap = phantomForest.getBlockMap();
uint_t sequenceId = blockSequenceRange.first;
for( auto it = blockMap.begin(); it != blockMap.end(); ++it )
mapping.insert( std::make_pair( it->first, sequenceId++ ) );
for( auto it = targetProcess.begin(); it != targetProcess.end(); ++it )
mapping.insert( std::make_pair( it->first->getId(), sequenceId++ ) );
WALBERLA_ASSERT_EQUAL( sequenceId, blockSequenceRange.second );
const std::vector<uint_t> neighborProcesses = phantomForest.getNeighboringProcesses();
......@@ -68,7 +75,11 @@ std::map< blockforest::BlockID, uint_t > getBlockIdToSequenceMapping( const Phan
mpi::BufferSystem bs( comm );
for( auto it = neighborProcesses.begin(); it != neighborProcesses.end(); ++it )
bs.sendBuffer( mpi::translateRank(mpi::MPIManager::instance()->comm(), comm, int_c(*it)) ) << mapping;
{
auto destRank = mpi::translateRank(mpi::MPIManager::instance()->comm(), comm, int_c(*it));
if (destRank != -1)
bs.sendBuffer( destRank ) << mapping;
}
bs.setReceiverInfoFromSendBufferState( false, true );
......@@ -101,6 +112,8 @@ T * ptr( std::vector<T> & v )
typedef uint_t idx_t;
bool DynamicParMetis::operator()( std::vector< std::pair< const PhantomBlock *, uint_t > > & targetProcess,
std::set< uint_t > & processesToRecvFrom,
const PhantomBlockForest & phantomForest,
......@@ -111,11 +124,11 @@ bool DynamicParMetis::operator()( std::vector< std::pair< const PhantomBlock *,
globalTimer.start();
//create new communicator which excludes processes which do not have blocks
MPI_Comm subComm;
MPI_Comm subComm = MPI_COMM_NULL;
MPI_Group allGroup, subGroup;
MPI_Comm_group( MPIManager::instance()->comm(), &allGroup );
std::vector<int> ranks;
if (phantomForest.getNumberOfBlocks() > 0)
if (targetProcess.size() > 0)
ranks.push_back( MPIManager::instance()->rank() );
ranks = mpi::allGatherv( ranks, MPIManager::instance()->comm() );
auto numSubProcesses = ranks.size();
......@@ -123,22 +136,51 @@ bool DynamicParMetis::operator()( std::vector< std::pair< const PhantomBlock *,
MPI_Group_incl(allGroup, int_c(ranks.size()), &ranks[0], &subGroup);
MPI_Comm_create( MPIManager::instance()->comm(), subGroup, &subComm);
if ( targetProcess.size() != 0)
{
int subRank;
int subSize;
MPI_Comm_rank(subComm, &subRank);
MPI_Comm_size(subComm, &subSize);
WALBERLA_CHECK_GREATER_EQUAL(subRank, 0);
WALBERLA_CHECK_LESS(subRank, subSize);
} else
{
int subRank;
MPI_Comm_rank(subComm, &subRank);
WALBERLA_CHECK_EQUAL(subRank, MPI_UNDEFINED);
}
int64_t edgecut = 0;
std::vector<int64_t> part( phantomForest.getNumberOfBlocks(), int64_c( MPIManager::instance()->rank() ) );
WALBERLA_CHECK_EQUAL( phantomForest.getNumberOfBlocks(), targetProcess.size() );
std::vector<int64_t> part( targetProcess.size(), int64_c( MPIManager::instance()->rank() ) );
if (subComm != MPI_COMM_NULL)
{
const std::pair<uint_t, uint_t> blockSequenceRange = getBlockSequenceRange( phantomForest, subComm );
const std::map< blockforest::BlockID, uint_t > mapping = getBlockIdToSequenceMapping( phantomForest, blockSequenceRange, subComm );
int subRank;
int subSize;
MPI_Comm_rank(subComm, &subRank);
MPI_Comm_size(subComm, &subSize);
WALBERLA_CHECK_UNEQUAL(targetProcess.size(), 0);
const std::pair<uint_t, uint_t> blockSequenceRange = getBlockSequenceRange( targetProcess.size(), subComm );
const std::map< blockforest::BlockID, uint_t > mapping = getBlockIdToSequenceMapping( phantomForest, targetProcess, blockSequenceRange, subComm ); //blockid to vertex id
std::vector<int64_t> vtxdist = mpi::allGather( int64_c( blockSequenceRange.second ), subComm );
vtxdist.insert( vtxdist.begin(), uint_t( 0 ) );
WALBERLA_CHECK_EQUAL( vtxdist.size(), subSize + 1 );
for (size_t i = 1; i < vtxdist.size(); ++i)
{
WALBERLA_ASSERT_LESS( vtxdist[i-1], vtxdist[i] );
}
std::vector<int64_t> adjncy, xadj, vsize, vwgt, adjwgt;
std::vector<double> xyz;
for( auto it = targetProcess.begin(); it != targetProcess.end(); ++it )
uint_t blockIndex = 0;
for( auto it = targetProcess.begin(); it != targetProcess.end(); ++it, ++blockIndex )
{
WALBERLA_CHECK_EQUAL(blockIndex, mapping.find(it->first->getId())->second - blockSequenceRange.first);
xadj.push_back( int64_c( adjncy.size() ) );
const PhantomBlock & block = *( it->first );
auto bi = block.getData< DynamicParMetisBlockInfo >();
......@@ -150,9 +192,12 @@ bool DynamicParMetis::operator()( std::vector< std::pair< const PhantomBlock *,
{
auto mapIt = mapping.find( nit->getId() );
WALBERLA_ASSERT_UNEQUAL( mapIt, mapping.end(), "BlockId of neighbor is not contained in sequence mapping!" );
WALBERLA_CHECK_GREATER_EQUAL( mapIt->second, 0 );
WALBERLA_CHECK_LESS( mapIt->second, vtxdist.back() );
adjncy.push_back( int64_c( mapIt->second ) );
auto edgeWeightIt = bi.getEdgeWeights().find( nit->getId() );
adjwgt.push_back( edgeWeightIt == bi.getEdgeWeights().end() ? int64_t( 0 ) : edgeWeightIt->second );
//WALBERLA_CHECK_UNEQUAL( edgeWeightIt->second, 0 ); // perhaps WARNING
adjwgt.push_back( edgeWeightIt == bi.getEdgeWeights().end() ? int64_t( 1 ) : edgeWeightIt->second );
}
break;
case PARMETIS_EDGES_FROM_EDGE_WEIGHTS:
......@@ -160,10 +205,15 @@ bool DynamicParMetis::operator()( std::vector< std::pair< const PhantomBlock *,
{
auto mapIt = mapping.find( edgeIt->first );
WALBERLA_ASSERT_UNEQUAL( mapIt, mapping.end(), "BlockId of neighbor is not contained in sequence mapping!" );
WALBERLA_CHECK_GREATER_EQUAL( mapIt->second, 0 );
WALBERLA_CHECK_LESS( mapIt->second, vtxdist.back() );
adjncy.push_back( int64_c( mapIt->second ) );
//WALBERLA_CHECK_UNEQUAL( edgeIt->second, 0 ); // perhaps WARNING
adjwgt.push_back( edgeIt->second );
}
break;
}
WALBERLA_CHECK_UNEQUAL( bi.getVertexWeight(), 0 );
vwgt.push_back( bi.getVertexWeight() );
vsize.push_back( bi.getVertexSize() );
xyz.push_back( bi.getVertexCoords()[0] );
......@@ -172,45 +222,58 @@ bool DynamicParMetis::operator()( std::vector< std::pair< const PhantomBlock *,
}
xadj.push_back( int64_c( adjncy.size() ) );
WALBERLA_ASSERT_EQUAL( vtxdist.size(), numSubProcesses + uint_t( 1 ) );
WALBERLA_ASSERT_EQUAL( xadj.size(), phantomForest.getNumberOfBlocks() + 1 );
WALBERLA_ASSERT_EQUAL( vwgt.size(), phantomForest.getNumberOfBlocks() );
WALBERLA_ASSERT_EQUAL( vsize.size(), phantomForest.getNumberOfBlocks() );
WALBERLA_ASSERT_EQUAL( adjncy.size(), adjwgt.size() );
int64_t wgtflag = weightsToUse_;
int64_t numflag = 0; // C-style ordering
int64_t ncon = 1; // Number of constraints
int64_t ndims = 3; // Number of dimensions
double ubvec[] = { real_t( 1.05 ) }; // imbalance tolerance
int64_t nparts = int64_c( MPIManager::instance()->numProcesses() ); // number of subdomains
double ipc2redist = real_t( 1000000.0 ); // compute repartitioning with low edge cut (set lower (down to 0.000001) to get minimal repartitioning )
MPI_Comm comm = subComm; //MPIManager::instance()->comm();
WALBERLA_CHECK_EQUAL( vtxdist.size(), numSubProcesses + uint_t( 1 ) );
WALBERLA_CHECK_EQUAL( xadj.size(), targetProcess.size() + 1 );
WALBERLA_CHECK_EQUAL( xadj.front(), 0);
WALBERLA_CHECK_EQUAL( xadj.back(), adjncy.size() );
for (size_t i = 1; i < xadj.size(); ++i)
{
WALBERLA_ASSERT_LESS( xadj[i-1], xadj[i] );
}
WALBERLA_CHECK_EQUAL( vwgt.size(), targetProcess.size() );
WALBERLA_CHECK_EQUAL( vsize.size(), targetProcess.size() );
WALBERLA_CHECK_EQUAL( xyz.size(), targetProcess.size() * 3 );
WALBERLA_CHECK_EQUAL( adjncy.size(), adjwgt.size() );
WALBERLA_CHECK_EQUAL( adjwgt.size(), xadj.back() );
int64_t wgtflag = weightsToUse_;
int64_t numflag = 0; // C-style ordering
int64_t ncon = 1; // Number of constraints
int64_t ndims = 3; // Number of dimensions
std::vector<double> ubvec( uint_c(ncon), double_c( 1.05 ) ); // imbalance tolerance
int64_t nparts = int64_c( MPIManager::instance()->numProcesses() ); // number of subdomains
double ipc2redist = double_c(ipc2redist_);
MPI_Comm comm = subComm; //MPIManager::instance()->comm();
std::vector<double> tpwgts( uint_c(nparts * ncon), 1.0 / double_c( nparts ) ); // vertex weight fraction that is stored in a subdomain
int64_t options[] = { int64_t( 1 ), int64_t( 0 ), int64_t( 23 ), int64_t( 1 ) };
std::vector<int64_t> options = { int64_t( 1 ), int64_t( 0 ), int64_t( 23 ), int64_t( 1 ) };
int metisResult = core::METIS_OK;
switch( algorithm_ )
{
case PARMETIS_PART_GEOM:
parmetisTimer.start();
metisResult = core::ParMETIS_V3_PartGeom( ptr( vtxdist ), &ndims, ptr( xyz ), ptr( part ), &comm );
parmetisTimer.end();
break;
case PARMETIS_PART_GEOM_KWAY:
parmetisTimer.start();
metisResult = core::ParMETIS_V3_PartGeomKway( ptr( vtxdist ), ptr( xadj ), ptr( adjncy ), ptr( vwgt ), ptr( adjwgt ), &wgtflag, &numflag, &ndims, ptr( xyz ), &ncon, &nparts, ptr( tpwgts ), ubvec, options, &edgecut, ptr( part ), &comm );
metisResult = core::ParMETIS_V3_PartGeomKway( ptr( vtxdist ), ptr( xadj ), ptr( adjncy ), ptr( vwgt ), ptr( adjwgt ), &wgtflag, &numflag, &ndims, ptr( xyz ), &ncon, &nparts, ptr( tpwgts ), ptr( ubvec ), ptr( options ), &edgecut, ptr( part ), &comm );
parmetisTimer.end();
break;
case PARMETIS_PART_KWAY:
parmetisTimer.start();
metisResult = core::ParMETIS_V3_PartKway( ptr( vtxdist ), ptr( xadj ), ptr( adjncy ), ptr( vwgt ), ptr( adjwgt ), &wgtflag, &numflag, &ncon, &nparts, ptr( tpwgts ), ubvec, options, &edgecut, ptr( part ), &comm );
metisResult = core::ParMETIS_V3_PartKway( ptr( vtxdist ), ptr( xadj ), ptr( adjncy ), ptr( vwgt ), ptr( adjwgt ), &wgtflag, &numflag, &ncon, &nparts, ptr( tpwgts ), ptr(ubvec), ptr(options), &edgecut, ptr( part ), &comm );
parmetisTimer.end();
break;
case PARMETIS_ADAPTIVE_REPART:
parmetisTimer.start();
metisResult = core::ParMETIS_V3_AdaptiveRepart( ptr( vtxdist ), ptr( xadj ), ptr( adjncy ), ptr( vwgt ), ptr( vsize ), ptr( adjwgt ), &wgtflag, &numflag, &ncon, &nparts, ptr( tpwgts ), ubvec, &ipc2redist, options, &edgecut, ptr( part ), &comm );
metisResult = core::ParMETIS_V3_AdaptiveRepart( ptr( vtxdist ), ptr( xadj ), ptr( adjncy ), ptr( vwgt ), ptr( vsize ), ptr( adjwgt ), &wgtflag, &numflag, &ncon, &nparts, ptr( tpwgts ), ptr(ubvec), &ipc2redist, ptr(options), &edgecut, ptr( part ), &comm );
parmetisTimer.end();
break;
case PARMETIS_REFINE_KWAY:
parmetisTimer.start();
metisResult = core::ParMETIS_V3_RefineKway( ptr( vtxdist ), ptr( xadj ), ptr( adjncy ), ptr( vwgt ), ptr( adjwgt ), &wgtflag, &numflag, &ncon, &nparts, ptr( tpwgts ), ubvec, options, &edgecut, ptr( part ), &comm );
metisResult = core::ParMETIS_V3_RefineKway( ptr( vtxdist ), ptr( xadj ), ptr( adjncy ), ptr( vwgt ), ptr( adjwgt ), &wgtflag, &numflag, &ncon, &nparts, ptr( tpwgts ), ptr(ubvec), ptr(options), &edgecut, ptr( part ), &comm );
parmetisTimer.end();
break;
}
......@@ -265,6 +328,8 @@ DynamicParMetis::Algorithm DynamicParMetis::stringToAlgorithm( std::string s )
if( s == "PART_GEOM_KWAY" )
return PARMETIS_PART_GEOM_KWAY;
else if( s == "PART_GEOM" )
return PARMETIS_PART_GEOM;
else if( s == "PART_KWAY" )
return PARMETIS_PART_KWAY;
else if( s == "PART_ADAPTIVE_REPART" )
......@@ -272,7 +337,7 @@ DynamicParMetis::Algorithm DynamicParMetis::stringToAlgorithm( std::string s )
else if( s == "REFINE_KWAY" )
return PARMETIS_REFINE_KWAY;
else
WALBERLA_ABORT( "Illegal ParMetis algorithm specified! Valid choices are: \"PART_GEOM_KWAY\", \"PART_KWAY\", \"PART_ADAPTIVE_REPART\", or \"REFINE_KWAY\"." );
WALBERLA_ABORT( "Illegal ParMetis algorithm specified (" << s << ")! Valid choices are: \"PART_GEOM_KWAY\", \"PART_KWAY\", \"PART_ADAPTIVE_REPART\", or \"REFINE_KWAY\"." );
}
......@@ -290,7 +355,7 @@ DynamicParMetis::WeightsToUse DynamicParMetis::stringToWeightsToUse( std::string
else if( s == "BOTH_WEIGHTS" )
return PARMETIS_BOTH_WEIGHTS;
else
WALBERLA_ABORT( "Illegal ParMetis weights usage specified! Valid choices are: \"NO_WEIGHTS\", \"EDGE_WEIGHTS\", \"VERTEX_WEIGHTS\", or \"BOTH_WEIGHTS\"." );
WALBERLA_ABORT( "Illegal ParMetis weights usage specified (" << s << ")! Valid choices are: \"NO_WEIGHTS\", \"EDGE_WEIGHTS\", \"VERTEX_WEIGHTS\", or \"BOTH_WEIGHTS\"." );
}
......@@ -304,7 +369,56 @@ DynamicParMetis::EdgeSource DynamicParMetis::stringToEdgeSource( std::string s )
else if( s == "EDGES_FROM_EDGE_WEIGHTS" )
return PARMETIS_EDGES_FROM_EDGE_WEIGHTS;
else
WALBERLA_ABORT( "Illegal ParMetis weights usage specified! Valid choices are: \"EDGES_FROM_FOREST\" or \"EDGES_FROM_EDGE_WEIGHTS\"" );
WALBERLA_ABORT( "Illegal ParMetis weights usage specified (" << s << ")! Valid choices are: \"EDGES_FROM_FOREST\" or \"EDGES_FROM_EDGE_WEIGHTS\"" );
}
std::string DynamicParMetis::algorithmToString( ) const
{
switch (algorithm_)
{
case walberla::blockforest::DynamicParMetis::PARMETIS_PART_GEOM_KWAY:
return "PART_GEOM_KWAY";
case walberla::blockforest::DynamicParMetis::PARMETIS_PART_GEOM:
return "PART_GEOM";
case walberla::blockforest::DynamicParMetis::PARMETIS_PART_KWAY:
return "PART_KWAY";
case walberla::blockforest::DynamicParMetis::PARMETIS_ADAPTIVE_REPART:
return "PART_ADAPTIVE_REPART";
case walberla::blockforest::DynamicParMetis::PARMETIS_REFINE_KWAY:
return "PARMETIS_REFINE_KWAY";
}
return "Unknown";
}
std::string DynamicParMetis::weightsToUseToString( ) const
{
switch (weightsToUse_)
{
case walberla::blockforest::DynamicParMetis::PARMETIS_NO_WEIGHTS:
return "NO_WEIGHTS";
case walberla::blockforest::DynamicParMetis::PARMETIS_EDGE_WEIGHTS:
return "EDGE_WEIGHTS";
case walberla::blockforest::DynamicParMetis::PARMETIS_VERTEX_WEIGHTS:
return "VERTEX_WEIGHTS";
case walberla::blockforest::DynamicParMetis::PARMETIS_BOTH_WEIGHTS:
return "BOTH_WEIGHTS";
}
return "Unknown";
}
std::string DynamicParMetis::edgeSourceToString( ) const
{
switch (edgeSource_)
{
case walberla::blockforest::DynamicParMetis::PARMETIS_EDGES_FROM_FOREST:
return "EDGES_FROM_FOREST";
case walberla::blockforest::DynamicParMetis::PARMETIS_EDGES_FROM_EDGE_WEIGHTS:
return "EDGES_FROM_EDGE_WEIGHTS";
}
return "Unknown";
}
......
......@@ -23,10 +23,12 @@
#include "blockforest/PhantomBlockForest.h"
#include "core/debug/Debug.h"
#include "core/DataTypes.h"
#include "core/math/Vector3.h"
#include "core/mpi/MPIWrapper.h"
#include <cmath>
#include <map>
namespace walberla {
......@@ -40,7 +42,7 @@ std::map< blockforest::BlockID, uint_t > getBlockIdToSequenceMapping( const Phan
class DynamicParMetis
{
public:
enum Algorithm { PARMETIS_PART_GEOM_KWAY, PARMETIS_PART_KWAY, PARMETIS_ADAPTIVE_REPART, PARMETIS_REFINE_KWAY };
enum Algorithm { PARMETIS_PART_GEOM, PARMETIS_PART_GEOM_KWAY, PARMETIS_PART_KWAY, PARMETIS_ADAPTIVE_REPART, PARMETIS_REFINE_KWAY };
enum WeightsToUse { PARMETIS_NO_WEIGHTS = 0, PARMETIS_EDGE_WEIGHTS = 1, PARMETIS_VERTEX_WEIGHTS = 2, PARMETIS_BOTH_WEIGHTS = 3 };
enum EdgeSource { PARMETIS_EDGES_FROM_FOREST, PARMETIS_EDGES_FROM_EDGE_WEIGHTS };
......@@ -54,6 +56,10 @@ public:
const PhantomBlockForest & phantomForest,
const uint_t iteration ) const;
void setipc2redist(double val) {ipc2redist_ = val;}
double getipc2redist() const {return ipc2redist_;}
bool edgeWeightsUsed() const { return ( weightsToUse_ == PARMETIS_EDGE_WEIGHTS ) || ( weightsToUse_ == PARMETIS_BOTH_WEIGHTS ); }
bool vertexWeightsUsed() const { return ( weightsToUse_ == PARMETIS_VERTEX_WEIGHTS ) || ( weightsToUse_ == PARMETIS_BOTH_WEIGHTS ); }
bool vertexSizeUsed() const { return algorithm_ == PARMETIS_ADAPTIVE_REPART; }
......@@ -62,10 +68,16 @@ public:
static WeightsToUse stringToWeightsToUse( std::string s );
static EdgeSource stringToEdgeSource( std::string s );
std::string algorithmToString() const;
std::string weightsToUseToString() const;
std::string edgeSourceToString() const;
protected:
Algorithm algorithm_;
WeightsToUse weightsToUse_;
EdgeSource edgeSource_;
double ipc2redist_ = real_t( 1000.0 ); ///< compute repartitioning with low edge cut (set lower (down to 0.000001) to get minimal repartitioning )
};
class DynamicParMetisBlockInfo
......@@ -90,7 +102,7 @@ public:
vsize_t getVertexSize() const { return vertexSize_; }
void setVertexSize( const vsize_t size ) { vertexSize_ = size; }
const Vector3<double> & getVertexCoords() const { return vertexCoords_; }
const Vector3<double> & getVertexCoords() const { WALBERLA_ASSERT( !std::isnan(vertexCoords_[0]) && !std::isnan(vertexCoords_[1]) && !std::isnan(vertexCoords_[2]) ); return vertexCoords_; }
void setVertexCoords( const Vector3<double> & p ) { vertexCoords_ = p; }
void setEdgeWeight( const blockforest::BlockID & blockId, const weight_t edgeWeight ) { edgeWeights_[blockId] = edgeWeight; }
......@@ -107,7 +119,7 @@ private:
weight_t vertexWeight_; /// Defines the weight of a block
vsize_t vertexSize_; /// Defines the cost of rebalancing a block
Vector3<double> vertexCoords_; /// Defines where the block is located in space. Needed by some ParMetis algorithms
Vector3<double> vertexCoords_ = Vector3<double>(std::numeric_limits<double>::signaling_NaN()); /// Defines where the block is located in space. Needed by some ParMetis algorithms
std::map< blockforest::BlockID, weight_t > edgeWeights_; /// Defines the cost of communication with other blocks
};
......
......@@ -20,28 +20,67 @@
#include "MPIHelper.h"
#include <core/debug/CheckFunctions.h>
namespace walberla {
namespace mpi {
//!
//! \brief This functions maps the rank in one communicator to the rank in another communicator.
//! \param srcComm source communicator
//! \param destComm destination communicator
//! \param srcRank rank in the source communicator
//! \return rank in the destination communicator or -1 if not available
//!
int translateRank(const MPI_Comm srcComm, const MPI_Comm destComm, const int srcRank)
{
if (srcComm == destComm)
{
return srcRank;
}
int destRank = -1;
MPI_Group srcGroup, destGroup;
MPI_Comm_group(srcComm, &srcGroup);
MPI_Comm_group(destComm, &destGroup);
MPI_Group_translate_ranks(srcGroup, 1, const_cast<int*>(&srcRank), destGroup, &destRank);
int size;
MPI_Comm_size(destComm, &size);
if (destRank == MPI_UNDEFINED) destRank = -1;
WALBERLA_CHECK_GREATER_EQUAL(destRank, -1);
WALBERLA_CHECK_LESS(destRank, size);
MPI_Group_free(&srcGroup);
MPI_Group_free(&destGroup);
return destRank;
}
//!
//! \brief This functions converts a array of ranks in one communicator to an array of ranks in another communicator.
//! \param srcComm source communicator
//! \param destComm destination communicator
//! \param srcRank source ranks
//! \return converted ranks, -1 if not available
//!
std::vector<int> translateRank(const MPI_Comm srcComm, const MPI_Comm destComm, const std::vector<int>& srcRank)
{
if (srcComm == destComm)
{
return srcRank;
}
std::vector<int> destRank(srcRank.size(), -1);
MPI_Group srcGroup, destGroup;
MPI_Comm_group(srcComm, &srcGroup);
MPI_Comm_group(destComm, &destGroup);
MPI_Group_translate_ranks(srcGroup, int_c(srcRank.size()), const_cast<int*>(&srcRank[0]), destGroup, &destRank[0]);
int size;
MPI_Comm_size(destComm, &size);
for (auto& dstRnk : destRank)
{
if (dstRnk == MPI_UNDEFINED) dstRnk = -1;
WALBERLA_CHECK_GREATER_EQUAL(dstRnk, -1);
WALBERLA_CHECK_LESS(dstRnk, size);
}
MPI_Group_free(&srcGroup);
MPI_Group_free(&destGroup);
return destRank;
......
......@@ -35,26 +35,38 @@ public:
typedef blockforest::DynamicParMetisBlockInfo PhantomBlockWeight;
typedef blockforest::DynamicParMetisBlockInfoPackUnpack PhantomBlockWeightPackUnpackFunctor;
MetisAssignmentFunctor( const shared_ptr<InfoCollection>& ic ) : ic_( ic )
{}
MetisAssignmentFunctor( shared_ptr<InfoCollection>& ic, const real_t baseWeight = real_t(10.0) ) : ic_(ic), baseWeight_(baseWeight) {}
void operator()( std::vector< std::pair< const PhantomBlock *, walberla::any > > & blockData, const PhantomBlockForest & )
{
for( auto it = blockData.begin(); it != blockData.end(); ++it )
{
const uint_t& weight = ic_->find( it->first->getId() )->second.numberOfLocalBodies;
blockforest::DynamicParMetisBlockInfo info( int64_c(weight) );
info.setVertexSize(int64_c( weight ));
const double weight = double_c( ic_->find( it->first->getId() )->second.numberOfLocalBodies ) + baseWeight_;
//const double commWeight = double_c( edgeWeightFactor * (double_c(ic_->find( it->first->getId() )->second.numberOfShadowBodies) + baseWeight_)) + 1;
blockforest::DynamicParMetisBlockInfo info( 0 );
info.setVertexWeight( int64_c(weight) );
info.setVertexSize( int64_c( weight ) );
info.setVertexCoords( it->first->getAABB().center() );
for( uint_t nb = uint_t(0); nb < it->first->getNeighborhoodSize(); ++nb )
{
info.setEdgeWeight(it->first->getNeighborId(nb), int64_c(weight) );
info.setEdgeWeight(it->first->getNeighborId(nb), int64_c(edgeWeight_) );
}
it->second = info;
}
}
inline void setBaseWeight( const double weight) { baseWeight_ = weight;}
inline double getBaseWeight() const { return baseWeight_; }
inline void setEdgeWeight( const double weight) { edgeWeight_ = weight;}
inline double getEdgeWeight() const { return edgeWeight_; }
private:
shared_ptr< InfoCollection > ic_;
///Base weight due to allocated data structures. A weight of zero for blocks is dangerous as empty blocks might accumulate on one process!
double baseWeight_ = 10.0;
double edgeWeight_ = 1.0;
};
}
......
......@@ -69,6 +69,11 @@ waLBerla_execute_test( NAME PE_OVERLAP )
waLBerla_compile_test( NAME PE_PARALLELEQUIVALENCE FILES ParallelEquivalence.cpp DEPENDS core blockforest )
waLBerla_execute_test( NAME PE_PARALLELEQUIVALENCE PROCESSES 4 )
if( WALBERLA_BUILD_WITH_PARMETIS )
waLBerla_compile_test( NAME PE_PARMETIS FILES ParMetis.cpp DEPENDS core blockforest )
waLBerla_execute_test( NAME PE_PARMETIS PROCESSES 64 )
endif()
waLBerla_compile_test( NAME PE_PARSEMESSAGE FILES ParseMessage.cpp DEPENDS core )
waLBerla_execute_test( NAME PE_PARSEMESSAGE )
......
//======================================================================================================================
//
// This file is part of waLBerla. waLBerla is free software: you can
// redistribute it and/or modify it under the terms of the GNU General Public
// License as published by the Free Software Foundation, either version 3 of
// the License, or (at your option) any later version.
//
// waLBerla is distributed in the hope that it will be useful, but WITHOUT
// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
// for more details.
//
// You should have received a copy of the GNU General Public License along
// with waLBerla (see COPYING.txt). If not, see <http://www.gnu.org/licenses/>.
//
//! \file ParMetis.cpp
//! \author Sebastian Eibl <sebastian.eibl@fau.de>
//
//======================================================================================================================
#include "pe/utility/CreateWorld.h"
#include <blockforest/loadbalancing/DynamicParMetis.h>
#include <core/Environment.h>
#include <core/logging/Logging.h>
#include <core/math/Sample.h>
using namespace walberla;
using namespace walberla::pe;
class ReGrid
{
public:
void operator()( std::vector< std::pair< const Block *, uint_t > > & minTargetLevels,
std::vector< const Block * > &, const BlockForest & /*forest*/ )
{
std::for_each( minTargetLevels.begin(),
minTargetLevels.end(),
[](auto& pair){pair.second = pair.first->getLevel() + 1;} );
}
};
class MetisAssignmentFunctor
{
public:
typedef blockforest::DynamicParMetisBlockInfo PhantomBlockWeight;
typedef blockforest::DynamicParMetisBlockInfoPackUnpack PhantomBlockWeightPackUnpackFunctor;
void operator()( std::vector< std::pair< const PhantomBlock *, walberla::any > > & blockData, const PhantomBlockForest & )
{
for( auto it = blockData.begin(); it != blockData.end(); ++it )
{
const auto& corner = it->first->getAABB().maxCorner();