openEMS/FDTD/openems_fdtd_mpi.cpp
Thorsten Liebig ab682cc0bd A few fixes for MPI
main.cpp:
	1. 	Check return value of ParseFDTDSetup and exit if false
	2. 	Use exit instead of return. These are almost identical. But
		in my OpenMPI installation the process with teh highes rank
		segfaults at the end when using return. This is not the case
		with exit. Probably some C++ cleanup problem (destructors).
openems.cpp:
	Give Parse_XML_FDTDSetup a deterministic return value.
openems_fdtd_mpi.cpp:
	1.	Remove the word "only" in an error message because there can
		also be too many processes.
	2.	Fix the indexing variables for SetSplitPos in SetupMPI. Otherwise
		more than one split results in an out-of-range exception and
		unexpected behavior.
RunOpenEMS_MPI.m:
	Apply Settings.MPI.GlobalArgs also to multi-host scenarios.
2016-12-02 19:03:35 +01:00

546 lines
15 KiB
C++

/*
* Copyright (C) 2011 Thorsten Liebig (Thorsten.Liebig@gmx.de)
*
* This program 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.
*
* This program 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 this program. If not, see <http://www.gnu.org/licenses/>.
*/
#include "openems_fdtd_mpi.h"
#include "FDTD/engine_interface_fdtd.h"
#include "FDTD/operator_mpi.h"
#include "FDTD/operator_cylinder.h"
#include "FDTD/engine_mpi.h"
#include "Common/processfields.h"
#include "Common/processintegral.h"
#include <stdio.h>
#include <stdlib.h>
#include <iostream>
#include <sstream>
#include <vector>
#include <iomanip>
#include <string.h>
#include <sys/time.h>
#include <time.h>
#include "mpi.h"
#include "tools/useful.h"
#include "tinyxml.h"
openEMS_FDTD_MPI::openEMS_FDTD_MPI(bool m_MPI_Debug) : openEMS()
{
m_MyID = MPI::COMM_WORLD.Get_rank();
m_NumProc = MPI::COMM_WORLD.Get_size();
m_MaxEnergy = 0;
m_EnergyDecrement = 1;
m_MPI_Op = NULL;
if (m_NumProc>1)
m_MPI_Enabled=true;
else
m_MPI_Enabled=false;
if (m_MyID==0)
{
m_Gather_Buffer = new int[m_NumProc];
m_Energy_Buffer = new double[m_NumProc];
}
else
{
m_Gather_Buffer = NULL;
m_Energy_Buffer = NULL;
}
m_MPI_Elem = NULL;
m_Original_Grid = NULL;
//redirect output to file for all ranks > 0
if ((m_MyID>0) && (m_MPI_Debug==false))
{
stringstream out_name;
out_name << "ID" << m_MyID << "_" << "output.txt";
m_Output = new ofstream();
m_Output->open(out_name.str().c_str());
cout.rdbuf(m_Output->rdbuf());
cerr.rdbuf(m_Output->rdbuf());
}
else
m_Output = NULL;
}
openEMS_FDTD_MPI::~openEMS_FDTD_MPI()
{
delete[] m_Gather_Buffer;
m_Gather_Buffer = NULL;
delete[] m_Energy_Buffer;
m_Energy_Buffer = NULL;
delete m_Original_Grid;
m_Original_Grid = NULL;
delete m_Output;
m_Output=NULL;
}
string openEMS_FDTD_MPI::GetExtLibsInfo()
{
stringstream str;
str << openEMS::GetExtLibsInfo();
// MPI
str << "\t\t" << "MPI -- Version: " << MPI_VERSION << "." << MPI_SUBVERSION << endl;
str << "\t\t" << " compiled against: ";
#ifdef MPICH2_VERSION
str << "MPICH2 " << MPICH2_VERSION << endl;
#endif
#ifdef OMPI_MAJOR_VERSION
str << "openMPI" << OMPI_MAJOR_VERSION << "." << OMPI_MINOR_VERSION << "." << OMPI_RELEASE_VERSION << endl;
#endif
return str.str();
}
bool openEMS_FDTD_MPI::parseCommandLineArgument( const char *argv )
{
if (!argv)
return false;
bool ret = openEMS::parseCommandLineArgument( argv );
if (ret)
return ret;
if (strcmp(argv,"--engine=MPI")==0)
{
cout << "openEMS_FDTD_MPI - enabled MPI parallel processing" << endl;
m_engine = EngineType_MPI;
return true;
}
return false;
}
bool openEMS_FDTD_MPI::Parse_XML_FDTDSetup(TiXmlElement* FDTD_Opts)
{
m_MPI_Elem = FDTD_Opts->FirstChildElement("MPI");
if (!m_MPI_Enabled)
{
if ((m_MPI_Elem!=NULL))
cerr << "openEMS_FDTD_MPI::SetupMPI: Warning: Number of MPI processes is 1, skipping MPI engine... " << endl;
return openEMS::Parse_XML_FDTDSetup(FDTD_Opts);
}
if (m_MPI_Elem==NULL)
{
MPI_Barrier(MPI_COMM_WORLD);
if (m_MyID==0)
cerr << "openEMS_FDTD_MPI::SetupMPI: Error: no MPI settings found, exiting MPI engine... " << endl;
exit(-1);
}
CSRectGrid* grid = m_CSX->GetGrid();
delete m_Original_Grid;
m_Original_Grid = CSRectGrid::Clone(grid);
string arg_Pos_Names[] = {"SplitPos_X", "SplitPos_Y", "SplitPos_Z"};
string arg_N_Names[] = {"SplitN_X", "SplitN_Y", "SplitN_Z"};
const char* tmp = NULL;
for (int n=0;n<3;++n)
{
m_SplitNumber[n].clear();
m_SplitNumber[n].push_back(0);
tmp = m_MPI_Elem->Attribute(arg_Pos_Names[n].c_str());
if (tmp) //check if a split position is requested
{
vector<double> SplitLines = SplitString2Double(tmp, ',');
bool inside;
unsigned int line;
for (size_t lineN = 0; lineN<SplitLines.size();++lineN)
{
line = m_Original_Grid->Snap2LineNumber(n, SplitLines.at(lineN), inside);
if (inside)
m_SplitNumber[n].push_back(line);
}
}
else //check if a number of splits is requested
{
int SplitN=0;
if (m_MPI_Elem->QueryIntAttribute( arg_N_Names[n].c_str(), &SplitN) == TIXML_SUCCESS)
{
if (SplitN>1)
{
vector<unsigned int> jobs = AssignJobs2Threads(m_Original_Grid->GetQtyLines(n)-1, SplitN, true);
unsigned int line=0;
for (size_t i = 0; i<jobs.size()-1;++i)
{
line += jobs.at(i);
m_SplitNumber[n].push_back(line);
}
}
}
}
m_SplitNumber[n].push_back(m_Original_Grid->GetQtyLines(n)-1);
unique(m_SplitNumber[n].begin(), m_SplitNumber[n].end());
}
return openEMS::Parse_XML_FDTDSetup(FDTD_Opts);
}
bool openEMS_FDTD_MPI::SetupMPI()
{
if (!m_MPI_Enabled)
return true;
MPI_Barrier(MPI_COMM_WORLD);
//validate number of processes
unsigned int numProcs = (m_SplitNumber[0].size()-1)*(m_SplitNumber[1].size()-1)*(m_SplitNumber[2].size()-1);
if (numProcs!=m_NumProc)
{
if (m_MyID==0)
cerr << "openEMS_FDTD_MPI::SetupMPI: Error: Requested splits require " << numProcs << " processes, but " << m_NumProc << " were found! Exit! " << endl;
exit(10);
}
//create process table
unsigned int procN = 0;
unsigned int splits[] = {(unsigned int)m_SplitNumber[0].size()-1, (unsigned int)m_SplitNumber[1].size()-1, (unsigned int)m_SplitNumber[2].size()-1};
m_MPI_Op->SetSplitNumbers(0,splits[0]);
m_MPI_Op->SetSplitNumbers(1,splits[1]);
m_MPI_Op->SetSplitNumbers(2,splits[2]);
unsigned int*** procTable=Create3DArray<unsigned int>(splits);
for (size_t i=0;i<m_SplitNumber[0].size()-1;++i)
for (size_t j=0;j<m_SplitNumber[1].size()-1;++j)
for (size_t k=0;k<m_SplitNumber[2].size()-1;++k)
{
procTable[i][j][k] = procN;
++procN;
}
m_MPI_Op->SetProcessTable(procTable);
CSRectGrid* grid = m_CSX->GetGrid();
//assign mesh and neighbors to this process
for (size_t i=0;i<m_SplitNumber[0].size()-1;++i)
{
for (size_t j=0;j<m_SplitNumber[1].size()-1;++j)
{
for (size_t k=0;k<m_SplitNumber[2].size()-1;++k)
{
if (procTable[i][j][k] == m_MyID)
{
m_MPI_Op->SetProcessTablePosition(0,i);
m_MPI_Op->SetProcessTablePosition(1,j);
m_MPI_Op->SetProcessTablePosition(2,k);
grid->ClearLines(0);
grid->ClearLines(1);
grid->ClearLines(2);
for (unsigned int n=m_SplitNumber[0].at(i);n<=m_SplitNumber[0].at(i+1);++n)
grid->AddDiscLine(0, m_Original_Grid->GetLine(0,n) );
for (unsigned int n=m_SplitNumber[1].at(j);n<=m_SplitNumber[1].at(j+1);++n)
grid->AddDiscLine(1, m_Original_Grid->GetLine(1,n) );
for (unsigned int n=m_SplitNumber[2].at(k);n<=m_SplitNumber[2].at(k+1);++n)
grid->AddDiscLine(2, m_Original_Grid->GetLine(2,n) );
m_MPI_Op->SetSplitPos(0,m_SplitNumber[0].at(i));
m_MPI_Op->SetSplitPos(1,m_SplitNumber[1].at(j));
m_MPI_Op->SetSplitPos(2,m_SplitNumber[2].at(k));
if (i>0)
m_MPI_Op->SetNeighborDown(0,procTable[i-1][j][k]);
if (i<m_SplitNumber[0].size()-2)
{
//add one additional line
grid->AddDiscLine(0, m_Original_Grid->GetLine(0,m_SplitNumber[0].at(i+1)+1 ));
m_MPI_Op->SetNeighborUp(0,procTable[i+1][j][k]);
}
if (j>0)
m_MPI_Op->SetNeighborDown(1,procTable[i][j-1][k]);
if (j<m_SplitNumber[1].size()-2)
{
//add one additional line
grid->AddDiscLine(1, m_Original_Grid->GetLine(1,m_SplitNumber[1].at(j+1)+1 ));
m_MPI_Op->SetNeighborUp(1,procTable[i][j+1][k]);
}
if (k>0)
m_MPI_Op->SetNeighborDown(2,procTable[i][j][k-1]);
if (k<m_SplitNumber[2].size()-2)
{
//add one additional line
grid->AddDiscLine(2, m_Original_Grid->GetLine(2,m_SplitNumber[2].at(k+1)+1 ));
m_MPI_Op->SetNeighborUp(2,procTable[i][j][k+1]);
}
}
}
}
}
m_MPI_Op->SetOriginalMesh(m_Original_Grid);
m_MPI_Op->SetTag(0);
return true;
}
bool openEMS_FDTD_MPI::SetupOperator()
{
bool ret = true;
if (m_engine == EngineType_MPI)
{
FDTD_Op = Operator_MPI::New();
}
else
{
ret = openEMS::SetupOperator();
}
m_MPI_Op = dynamic_cast<Operator_MPI*>(FDTD_Op);
if ((m_MPI_Enabled) && (m_MPI_Op==NULL))
{
cerr << "openEMS_FDTD_MPI::SetupOperator: Error: MPI is enabled but requested engine does not support MPI... EXIT!!!" << endl;
MPI_Barrier(MPI_COMM_WORLD);
exit(0);
}
ret &= SetupMPI();
return ret;
}
unsigned int openEMS_FDTD_MPI::GetNextStep()
{
//start processing and get local next step
int step=PA->Process();
double currTS = FDTD_Eng->GetNumberOfTimesteps();
if ((step<0) || (step>(int)(NrTS - currTS))) step=NrTS - currTS;
int local_step=step;
//find the smallest next step requestes by all processings
MPI_Reduce(&local_step, &step, 1, MPI_INT, MPI_MIN, 0, MPI_COMM_WORLD);
//send the smallest next step to all
MPI_Bcast(&step, 1, MPI_INT, 0, MPI_COMM_WORLD);
return step;
}
bool openEMS_FDTD_MPI::CheckEnergyCalc()
{
int local_Check = (int)m_ProcField->CheckTimestep();
int result;
//check if some process request an energy calculation --> the sum is larger than 0
MPI_Reduce(&local_Check, &result, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
//send result to all
MPI_Bcast(&result, 1, MPI_INT, 0, MPI_COMM_WORLD);
//calc energy if result is non-zero
return result>0;
}
double openEMS_FDTD_MPI::CalcEnergy()
{
double energy = 0;
double loc_energy= m_ProcField->CalcTotalEnergyEstimate();
//calc the sum of all local energies
MPI_Reduce(&loc_energy, &energy, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
//send sum-energy to all processes
MPI_Bcast(&energy, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
if (energy>m_MaxEnergy)
m_MaxEnergy = energy;
if (m_MaxEnergy)
m_EnergyDecrement = energy/m_MaxEnergy;
return energy;
}
bool openEMS_FDTD_MPI::SetupProcessing()
{
bool ret = openEMS::SetupProcessing();
//search for active processings in different processes
size_t numProc = PA->GetNumberOfProcessings();
int active=0;
bool deactivate = false;
bool rename = false;
for (size_t n=0;n<numProc;++n)
{
Processing* proc = PA->GetProcessing(n);
int isActive = (int)proc->GetEnable();
//sum of all active processings
MPI_Reduce(&isActive, &active, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
deactivate = false;
rename = false;
if ((m_MyID==0) && (active>1)) //more than one active processing...
{
deactivate = true; //default
if (dynamic_cast<ProcessIntegral*>(proc)!=NULL)
{
//type is integral processing --> disable! Needs to be fixed!
cerr << "openEMS_FDTD_MPI::SetupProcessing(): Warning: Processing: " << proc->GetName() << " occures multiple times and is being deactivated..." << endl;
deactivate = true;
rename = false;
}
if (dynamic_cast<ProcessFields*>(proc)!=NULL)
{
//type is field processing --> renameing! Needs to be fixed!
cerr << "openEMS_FDTD_MPI::SetupProcessing(): Warning: Processing: " << proc->GetName() << " occures multiple times and is being renamed..." << endl;
deactivate = false;
rename = true;
}
}
//broadcast informations to all
MPI_Bcast(&deactivate, 1, MPI::BOOL, 0, MPI_COMM_WORLD);
MPI_Bcast(&rename, 1, MPI::BOOL, 0, MPI_COMM_WORLD);
if (deactivate)
proc->SetEnable(false);
if (rename)
{
ProcessFields* ProcField = dynamic_cast<ProcessFields*>(proc);
if (ProcField)
{
stringstream name_ss;
name_ss << "ID" << m_MyID << "_" << ProcField->GetName();
ProcField->SetName(name_ss.str());
ProcField->SetFileName(name_ss.str());
}
}
}
return ret;
}
int openEMS_FDTD_MPI::SetupFDTD()
{
return openEMS::SetupFDTD();
}
void openEMS_FDTD_MPI::RunFDTD()
{
if (!m_MPI_Enabled)
return openEMS::RunFDTD();
cout << "Running MPI-FDTD engine... this may take a while... grab a cup of coffee?!?" << endl;
//get the sum of all cells
unsigned int local_NrCells=FDTD_Op->GetNumberCells();
MPI_Reduce(&local_NrCells, &m_NumberCells, 1, MPI_UNSIGNED, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Bcast(&m_NumberCells, 1, MPI_UNSIGNED, 0, MPI_COMM_WORLD);
//special handling of a field processing, needed to realize the end criteria...
m_ProcField = new ProcessFields(NewEngineInterface());
PA->AddProcessing(m_ProcField);
//init processings
PA->InitAll();
double currE=0;
//add all timesteps to end-crit field processing with max excite amplitude
unsigned int maxExcite = FDTD_Op->GetExcitationSignal()->GetMaxExcitationTimestep();
// for (unsigned int n=0; n<FDTD_Op->Exc->Volt_Count; ++n)
// m_ProcField->AddStep(FDTD_Op->Exc->Volt_delay[n]+maxExcite);
m_ProcField->AddStep(maxExcite);
int prevTS=0,currTS=0;
double numCells = FDTD_Op->GetNumberCells();
double speed = 0;
double t_diff;
double t_run;
timeval currTime;
gettimeofday(&currTime,NULL);
timeval startTime = currTime;
timeval prevTime= currTime;
if (m_DumpStats)
InitRunStatistics(__OPENEMS_RUN_STAT_FILE__);
//*************** simulate ************//
PA->PreProcess();
int step = GetNextStep();
while ((step>0) && !CheckAbortCond())
{
FDTD_Eng->IterateTS(step);
step = GetNextStep();
currTS = FDTD_Eng->GetNumberOfTimesteps();
currE = 0;
gettimeofday(&currTime,NULL);
t_diff = CalcDiffTime(currTime,prevTime);
if (CheckEnergyCalc())
currE = CalcEnergy();
//make sure all processes are at the same simulation time
MPI_Bcast(&t_diff, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
if (t_diff>4)
{
if (currE==0)
currE = CalcEnergy();
if (m_MyID==0)
{
t_run = CalcDiffTime(currTime,startTime);
speed = numCells*(currTS-prevTS)/t_diff;
cout << "[@" << FormatTime(t_run) << "] Timestep: " << setw(12) << currTS ;
cout << " || Speed: " << setw(6) << setprecision(1) << std::fixed << speed*1e-6 << " MC/s (" << setw(4) << setprecision(3) << std::scientific << t_diff/(currTS-prevTS) << " s/TS)" ;
cout << " || Energy: ~" << setw(6) << setprecision(2) << std::scientific << currE << " (-" << setw(5) << setprecision(2) << std::fixed << fabs(10.0*log10(m_EnergyDecrement)) << "dB)" << endl;
//set step to zero to abort simulation and send to all
if (m_EnergyDecrement<endCrit)
step=0;
if (m_DumpStats)
DumpRunStatistics(__OPENEMS_RUN_STAT_FILE__, t_run, currTS, speed, currE);
}
MPI_Bcast(&step, 1, MPI_INT, 0, MPI_COMM_WORLD);
prevTime=currTime;
prevTS=currTS;
PA->FlushNext();
}
}
if ((m_MyID==0) && (m_EnergyDecrement>endCrit) && (FDTD_Op->GetExcitationSignal()->GetExciteType()==0))
cerr << "RunFDTD: max. number of timesteps was reached before the end-criteria of -" << fabs(10.0*log10(endCrit)) << "dB was reached... " << endl << \
"\tYou may want to choose a higher number of max. timesteps... " << endl;
gettimeofday(&currTime,NULL);
t_diff = CalcDiffTime(currTime,startTime);
if (m_MyID==0)
{
cout << "Time for " << FDTD_Eng->GetNumberOfTimesteps() << " iterations with " << FDTD_Op->GetNumberCells() << " cells : " << t_diff << " sec" << endl;
cout << "Speed: " << numCells*(double)FDTD_Eng->GetNumberOfTimesteps()/t_diff*1e-6 << " MCells/s " << endl;
if (m_DumpStats)
DumpStatistics(__OPENEMS_STAT_FILE__, t_diff);
}
//*************** postproc ************//
PA->PostProcess();
}