G4MPImanager.cc

Go to the documentation of this file.

//
// ********************************************************************
// * License and Disclaimer                                           *
// *                                                                  *
// * The  Geant4 software  is  copyright of the Copyright Holders  of *
// * the Geant4 Collaboration.  It is provided  under  the terms  and *
// * conditions of the Geant4 Software License,  included in the file *
// * LICENSE and available at  http://cern.ch/geant4/license .  These *
// * include a list of copyright holders.                             *
// *                                                                  *
// * Neither the authors of this software system, nor their employing *
// * institutes,nor the agencies providing financial support for this *
// * work  make  any representation or  warranty, express or implied, *
// * regarding  this  software system or assume any liability for its *
// * use.  Please see the license in the file  LICENSE  and URL above *
// * for the full disclaimer and the limitation of liability.         *
// *                                                                  *
// * This  code  implementation is the result of  the  scientific and *
// * technical work of the GEANT4 collaboration.                      *
// * By using,  copying,  modifying or  distributing the software (or *
// * any work based  on the software)  you  agree  to acknowledge its *
// * use  in  resulting  scientific  publications,  and indicate your *
// * acceptance of all terms of the Geant4 Software license.          *
// ********************************************************************
/// @file G4MPImanager.cc
/// @brief MPI manager class
 
#include "mpi.h"
#include <getopt.h>
#include <stdio.h>
#include <time.h>
#include "G4Run.hh"
#include "G4RunManager.hh"
#include "G4StateManager.hh"
#include "G4UIcommand.hh"
#include "G4UImanager.hh"
#include "G4MPIbatch.hh"
#include "G4MPImanager.hh"
#include "G4MPImessenger.hh"
#include "G4MPIrandomSeedGenerator.hh"
#include "G4MPIsession.hh"
#include "G4MPIstatus.hh"
#include "G4MPIextraWorker.hh"
 
G4MPImanager* G4MPImanager::g4mpi_ = NULL;
 
// --------------------------------------------------------------------------
namespace {
 
// wrappers for thread functions
void thread_ExecuteThreadCommand(const G4String* command)
{
  G4MPImanager::GetManager()-> ExecuteThreadCommand(*command);
}
 
// --------------------------------------------------------------------------
void Wait(G4int ausec)
{
  struct timespec treq, trem;
  treq.tv_sec = 0;
  treq.tv_nsec = ausec*1000;
 
  nanosleep(&treq, &trem);
}
 
} // end of namespace
 
// --------------------------------------------------------------------------
G4MPImanager::G4MPImanager(int nof_extra_workers)
  : verbose_(0), 
    COMM_G4COMMAND_(MPI_COMM_NULL), processing_comm_(MPI_COMM_NULL), 
    collecting_comm_(MPI_COMM_NULL), all_comm_(MPI_COMM_NULL),
    qfcout_(false), qinitmacro_(false), qbatchmode_(false),
    thread_id_(0), master_weight_(1.), nof_extra_workers_(nof_extra_workers)
{
  //MPI::Init();
  MPI::Init_thread(MPI::THREAD_SERIALIZED);
  Initialize();
}
 
// --------------------------------------------------------------------------
G4MPImanager::G4MPImanager(int argc, char** argv, int nof_extra_workers)
  : verbose_(0), 
    COMM_G4COMMAND_(MPI_COMM_NULL), processing_comm_(MPI_COMM_NULL), 
    collecting_comm_(MPI_COMM_NULL), all_comm_(MPI_COMM_NULL),
    qfcout_(false), qinitmacro_(false), qbatchmode_(false),
    thread_id_(0), master_weight_(1.), nof_extra_workers_(nof_extra_workers)
{
  //MPI::Init(argc, argv);
  MPI::Init_thread(argc, argv, MPI::THREAD_SERIALIZED);
  Initialize();
  ParseArguments(argc, argv);
}
 
// --------------------------------------------------------------------------
G4MPImanager::~G4MPImanager()
{
  if( is_slave_ && qfcout_ ) fscout_.close();
 
  delete status_;
  delete messenger_;
  delete session_;
 
  if ( nof_extra_workers_ ) {
    MPI_Group_free(&world_group_);
    MPI_Group_free(&processing_group_);
    MPI_Group_free(&collecting_group_);
    MPI_Group_free(&all_group_);
    if (processing_comm_ != MPI_COMM_NULL) {
      MPI_Comm_free(&processing_comm_);
    }
    if (collecting_comm_ != MPI_COMM_NULL) {
      MPI_Comm_free(&collecting_comm_);
    }
    if (all_comm_ != MPI_COMM_NULL) {
      MPI_Comm_free(&all_comm_);
    }
  }
  else {
    COMM_G4COMMAND_.Free();
  }
 
  MPI::Finalize();
}
 
// --------------------------------------------------------------------------
G4MPImanager* G4MPImanager::GetManager()
{
  if ( g4mpi_ == NULL ) {
    G4Exception("G4MPImanager::GetManager()", "MPI001",
                FatalException, "G4MPImanager is not instantiated.");
  }
  return g4mpi_;
}
 
// --------------------------------------------------------------------------
void G4MPImanager::SetExtraWorker(G4VMPIextraWorker* extraWorker)
{
  if ( ! nof_extra_workers_ ) {
    G4Exception("G4MPImanager::SetExtraWorker()", "MPI001",
      FatalException, "Number of extra workers >0 must be set first.");
  }
 
  extra_worker_ = extraWorker;
}
 
// --------------------------------------------------------------------------
void G4MPImanager::Initialize()
{
  // G4cout << "G4MPImanager::Initialize" << G4endl;
 
  if ( g4mpi_ != NULL ) {
    G4Exception("G4MPImanager::Initialize()", "MPI002",
                FatalException, "G4MPImanager is already instantiated.");
  }
 
  g4mpi_ = this;
 
  // get rank information
  world_size_ = MPI::COMM_WORLD.Get_size();
  if ( world_size_ - nof_extra_workers_ <= 0 ) {
    G4Exception("G4MPImanager::SetExtraWorker()", "MPI001",
    JustWarning, "Cannot reserve extra ranks: the MPI size is not sufficient.");
    nof_extra_workers_ = 0;
  }
  size_ = world_size_ - nof_extra_workers_;
  rank_ = MPI::COMM_WORLD.Get_rank();
  is_master_ = (rank_ == kRANK_MASTER);
  is_slave_ = (rank_ != kRANK_MASTER);
  is_extra_worker_ = false;
 
  if ( nof_extra_workers_ ) {
    // G4cout << "Extra workers requested" << G4endl;
 
    // Define three groups of workers: processing, collecting and all;
    // if no extra workers are declared, all world ranks are processing ranks
 
    // MPI_Group world_group;
    MPI_Comm_group(MPI_COMM_WORLD, &world_group_);
    
    // Group 1 - processing ranks
    int* ranks1 =  new int[size_];
    for (int i=0; i<size_; i++) ranks1[i] = i;
    // Construct a group containing all of the processing ranks in world_group
    MPI_Group_incl(world_group_, size_, ranks1, &processing_group_);
    
    // Group 2 - collecting ranks
    int* ranks2 =  new int[nof_extra_workers_];
    for (int i=0; i<nof_extra_workers_; i++) ranks2[i] = (world_size_ - nof_extra_workers_)  + i;
    // Construct a group containing all of the collecting ranks in world_group
    MPI_Group_incl(world_group_, nof_extra_workers_, ranks2, &collecting_group_);
    
    // Group 3 - all ranks
    int* ranks3 = new int[world_size_];
    for (int i=0; i<world_size_; i++) ranks3[i] = i;
    // Construct a group containing all of the processing ranks in world_group
    MPI_Group_incl(world_group_, world_size_, ranks3, &all_group_);
 
    // Create new communicators based on the groups
    MPI_Comm_create_group(MPI_COMM_WORLD, processing_group_, 0, &processing_comm_);
    MPI_Comm_create_group(MPI_COMM_WORLD, collecting_group_, 0, &collecting_comm_);
    MPI_Comm_create_group(MPI_COMM_WORLD, all_group_, 0, &all_comm_);
 
    // COMM_G4COMMAND_ = processing_comm_ copy
    COMM_G4COMMAND_ = MPI::Intracomm(processing_comm_);
 
  } else {
    // G4cout << "No extra workers requested" << G4endl;
    // initialize MPI communicator
    COMM_G4COMMAND_ = MPI::COMM_WORLD.Dup();
  }
 
  is_extra_worker_ = (collecting_comm_ != MPI_COMM_NULL);
 
  // new G4MPI stuffs
  messenger_ = new G4MPImessenger();
  messenger_-> SetTargetObject(this);
  session_ = new G4MPIsession;
  status_ = new G4MPIstatus;
 
  if ( ! is_extra_worker_ ) {
    // default seed generator is random generator.
    seed_generator_ = new G4MPIrandomSeedGenerator;
    DistributeSeeds();
  }
 
  // print status of this worker
  // G4cout << this << " world_size_ " << world_size_ << G4endl;
  // G4cout << this << " size_ " << size_ << G4endl;
  // G4cout << this << " nof_extra_workers_  " << nof_extra_workers_  << G4endl;
  // G4cout << this << " is_master_ " << is_master_ << G4endl;
  // G4cout << this << " is_slave_ " << is_slave_ << G4endl;
  // G4cout << this << " is_extra_worker_ " << is_extra_worker_ << G4endl;
  // G4cout << this << " is_processing_worker_ " 
  //        <<  (processing_comm_ != MPI_COMM_NULL) << G4endl;
}
 
// --------------------------------------------------------------------------
void G4MPImanager::ParseArguments(int argc, char** argv)
{
  G4int qhelp = 0;
  G4String ofprefix = "mpi";
 
  G4int c;
  while ( 1 ) {
    G4int option_index = 0;
    static struct option long_options[] = {
      {"help",    no_argument,        NULL, 'h'},
      {"verbose", no_argument,        NULL, 'v'},
      {"init",    required_argument,  NULL, 'i'},
      {"ofile",   optional_argument,  NULL, 'o'},
      {NULL,      0,                  NULL,  0}
    };
 
    opterr = 0; // suppress message
    c = getopt_long(argc, argv, "hvi:o", long_options, &option_index);
    opterr = 1;
 
    if( c == -1 ) break;
 
    switch (c) {
    case 'h' :
      qhelp = 1;
      break;
    case 'v' :
      verbose_ = 1;
      break;
    case 'i' :
      qinitmacro_ = true;
      init_file_name_ = optarg;
      break;
    case 'o' :
      qfcout_ = true;
      if ( optarg ) ofprefix = optarg;
      break;
    default:
      G4cerr << "*** invalid options specified." << G4endl;
      std::exit(EXIT_FAILURE);
      break;
    }
  }
 
  // show help
  if ( qhelp ) {
    if ( is_master_ ) ShowHelp();
    MPI::Finalize();
    std::exit(EXIT_SUCCESS);
  }
 
  // file output
  if( is_slave_ && qfcout_ ) {
    G4String prefix = ofprefix + ".%03d" + ".cout";
    char str[1024];
    sprintf(str, prefix.c_str(), rank_);
    G4String fname(str);
    fscout_.open(fname.c_str(), std::ios::out);
  }
 
  // non-option ARGV-elements ...
  if ( optind < argc ) {
    qbatchmode_ = true;
    macro_file_name_ = argv[optind];
  }
}
 
// ====================================================================
void G4MPImanager::UpdateStatus()
{
  G4RunManager* runManager = G4RunManager::GetRunManager();
  const G4Run* run = runManager-> GetCurrentRun();
 
  G4int runid, eventid, neventTBP;
 
  G4StateManager* stateManager = G4StateManager::GetStateManager();
  G4ApplicationState g4state = stateManager-> GetCurrentState();
 
  if ( run ) {
    runid = run-> GetRunID();
    neventTBP = run -> GetNumberOfEventToBeProcessed();
    eventid = run-> GetNumberOfEvent();
    if( g4state == G4State_GeomClosed || g4state == G4State_EventProc ) {
      status_-> StopTimer();
    }
  } else {
    runid = 0;
    eventid = 0;
    neventTBP = 0;
  }
 
  status_-> SetStatus(rank_, runid, neventTBP, eventid, g4state);
}
 
// --------------------------------------------------------------------------
void G4MPImanager::ShowStatus()
{
  G4int buff[G4MPIstatus::kNSIZE];
 
  UpdateStatus();
  G4bool gstatus = CheckThreadStatus();
 
  if ( is_master_ ) {
    status_-> Print();  // for maser itself
 
    G4int nev = status_-> GetEventID();
    G4int nevtp = status_-> GetNEventToBeProcessed();
    G4double cputime = status_-> GetCPUTime();
 
    // receive from each slave
    for ( G4int islave = 1; islave < size_; islave++ ) {
      COMM_G4COMMAND_.Recv(buff, G4MPIstatus::kNSIZE, MPI::INT,
                          islave, kTAG_G4STATUS);
      status_-> UnPack(buff);
      status_-> Print();
 
      // aggregation
      nev += status_-> GetEventID();
      nevtp += status_-> GetNEventToBeProcessed();
      cputime += status_-> GetCPUTime();
    }
 
    G4String strStatus;
    if ( gstatus ) {
      strStatus = "Run";
    } else {
      strStatus = "Idle";
    }
 
    G4cout << "-------------------------------------------------------"
           << G4endl
           << "* #ranks= " << size_
           << "   event= " << nev << "/" << nevtp
           << " state= " << strStatus
           << " time= " << cputime << "s"
           << G4endl;
  } else {
    status_-> Pack(buff);
    COMM_G4COMMAND_.Send(buff, G4MPIstatus::kNSIZE, MPI::INT,
                         kRANK_MASTER, kTAG_G4STATUS);
  }
}
 
// ====================================================================
void G4MPImanager::DistributeSeeds()
{
  // Do nothing if not processing worker
  if ( is_extra_worker_ ) return;
 
  std::vector<G4long> seed_list = seed_generator_-> GetSeedList();
  G4Random::setTheSeed(seed_list[rank_]);
}
 
// --------------------------------------------------------------------------
void G4MPImanager::ShowSeeds()
{
  G4long buff;
 
  if ( is_master_ ) {
    // print master
    G4cout << "* rank= " << rank_
           << " seed= " << G4Random::getTheSeed()
           << G4endl;
    // receive from each slave
    for ( G4int islave = 1; islave < size_; islave++ ) {
      COMM_G4COMMAND_.Recv(&buff, 1, MPI::LONG, islave, kTAG_G4SEED);
      G4cout << "* rank= " << islave
             << " seed= " << buff
             << G4endl;
    }
  } else { // slaves
    buff = G4Random::getTheSeed();
    COMM_G4COMMAND_.Send(&buff, 1, MPI::LONG, kRANK_MASTER, kTAG_G4SEED);
  }
}
 
// --------------------------------------------------------------------------
void G4MPImanager::SetSeed(G4int inode, G4long seed)
{
  if( rank_ == inode ) {
    CLHEP::HepRandom::setTheSeed(seed);
  }
}
 
// ====================================================================
G4bool G4MPImanager::CheckThreadStatus()
{
  unsigned buff;
  unsigned qstatus = 0;
 
  if( is_master_ ) {
    qstatus = (thread_id_ != 0);
    // get slave status
    for ( G4int islave = 1; islave < size_; islave++ ) {
      MPI::Request request = COMM_G4COMMAND_.Irecv(&buff, 1, MPI::UNSIGNED,
                                                  islave, kTAG_G4STATUS);
      while( ! request.Test() ) {
        ::Wait(1000);
      }
      qstatus |= buff;
    }
  } else {
    buff = (thread_id_ !=0);
    COMM_G4COMMAND_.Send(&buff, 1, MPI::UNSIGNED, kRANK_MASTER, kTAG_G4STATUS);
  }
 
  // broadcast
  buff = qstatus; // for master
  COMM_G4COMMAND_.Bcast(&buff, 1, MPI::UNSIGNED, kRANK_MASTER);
  qstatus = buff; // for slave
 
  if ( qstatus != 0 ) return true;
  else return false;
}
 
// --------------------------------------------------------------------------
void G4MPImanager::ExecuteThreadCommand(const G4String& command)
{
  // this method is a thread function.
  G4UImanager* UI = G4UImanager::GetUIpointer();
  G4int rc = UI-> ApplyCommand(command);
 
  G4int commandStatus = rc - (rc%100);
 
  switch( commandStatus ) {
  case fCommandSucceeded:
    break;
  case fIllegalApplicationState:
    G4cerr << "illegal application state -- command refused" << G4endl;
    break;
  default:
    G4cerr << "command refused (" << commandStatus << ")" << G4endl;
    break;
  }
 
  // thread is joined
  if ( thread_id_ ) {
    pthread_join(thread_id_, 0);
    thread_id_ = 0;
  }
 
  return;
}
 
// --------------------------------------------------------------------------
void G4MPImanager::ExecuteBeamOnThread(const G4String& command)
{
  G4bool threadStatus = CheckThreadStatus();
 
  if (threadStatus) {
    if ( is_master_ ) {
      G4cout << "G4MPIsession:: beamOn is still running." << G4endl;
    }
  } else { // ok
    static G4String cmdstr;
    cmdstr = command;
    G4int rc = pthread_create(&thread_id_, 0,
                              (Func_t)thread_ExecuteThreadCommand,
                              (void*)&cmdstr);
    if (rc != 0)
      G4Exception("G4MPImanager::ExecuteBeamOnThread()",
                  "MPI003", FatalException,
                  "Failed to create a beamOn thread.");
  }
}
 
// --------------------------------------------------------------------------
void G4MPImanager::JoinBeamOnThread()
{
  if ( thread_id_ ) {
    pthread_join(thread_id_, 0);
    thread_id_ = 0;
  }
}
 
// ====================================================================
G4String G4MPImanager::BcastCommand(const G4String& command)
{
  // Do nothing if not processing worker
  if (is_extra_worker_) return G4String("exit");
 
  enum { kBUFF_SIZE = 512 };
  static char sbuff[kBUFF_SIZE];
  command.copy(sbuff, kBUFF_SIZE);
  G4int len = command.size();
  sbuff[len] ='\0';   // no boundary check
 
  // "command" is not yet fixed in slaves at this time.
 
  // waiting message exhausts CPU in LAM!
  //COMM_G4COMMAND_.Bcast(sbuff, ssize, MPI::CHAR, RANK_MASTER);
 
  // another implementation
  if( is_master_ ) {
    for ( G4int islave = 1; islave < size_; islave++ ) {
      COMM_G4COMMAND_.Send(sbuff, kBUFF_SIZE, MPI::CHAR,
                           islave, kTAG_G4COMMAND);
    }
  } else {
    // try non-blocking receive
    MPI::Request request= COMM_G4COMMAND_.Irecv(sbuff, kBUFF_SIZE, MPI::CHAR,
                                                kRANK_MASTER, kTAG_G4COMMAND);
    // polling...
    while(! request.Test()) {
      ::Wait(1000);
    }
  }
 
  return G4String(sbuff);
}
 
// ====================================================================
void G4MPImanager::ExecuteMacroFile(const G4String& fname, G4bool qbatch)
{
  G4bool currentmode = qbatchmode_;
  qbatchmode_ = true;
  G4MPIbatch* batchSession = new G4MPIbatch(fname, qbatch);
  batchSession-> SessionStart();
  delete batchSession;
  qbatchmode_ = currentmode;
}
 
// --------------------------------------------------------------------------
void G4MPImanager::BeamOn(G4int nevent, G4bool qdivide)
{
  // G4cout << "G4MPImanager::BeamOn " << nevent << G4endl;
 
#ifndef G4MULTITHREADED
  G4RunManager* runManager = G4RunManager::GetRunManager();
#endif
 
  if ( qdivide ) { // events are divided
    G4double ntot = master_weight_ + size_ - 1.;
    G4int nproc = G4int(nevent/ntot);
    G4int nproc0 = nevent - nproc*(size_ - 1);
    fevents_in_master = nproc0;
    fevents_in_slave = nproc;
 
    if ( verbose_ > 0 && is_master_ ) {
      G4cout << "#events in master=" << nproc0 << " / "
             << "#events in slave="  << nproc << G4endl;
    }
 
    status_-> StartTimer(); // start timer
 
#ifdef G4MULTITHREADED
    G4String str_nevt;
    if ( is_master_ ) str_nevt = G4UIcommand::ConvertToString(nproc0);
    else str_nevt = G4UIcommand::ConvertToString(nproc);
    G4UImanager* UI = G4UImanager::GetUIpointer();
    UI-> ApplyCommand("/run/beamOn " + str_nevt);
#else
    if ( is_master_ ) runManager-> BeamOn(nproc0);
    else runManager-> BeamOn(nproc);
#endif
 
    status_-> StopTimer(); // stop timer
 
  } else { // same events are generated in each node (for test use)
    if( verbose_ > 0 && is_master_ ) {
      G4cout << "#events in master=" << nevent << " / "
             << "#events in slave="  << nevent << G4endl;
    }
    status_-> StartTimer(); // start timer
 
#ifdef G4MULTITHREADED
    G4String str_nevt = G4UIcommand::ConvertToString(nevent);
    G4UImanager* UI = G4UImanager::GetUIpointer();
    UI-> ApplyCommand("/run/beamOn " + str_nevt);
#else
    runManager-> BeamOn(nevent);
#endif
 
    status_-> StopTimer(); // stop timer
  }
}
 
// --------------------------------------------------------------------------
void G4MPImanager::WaitBeamOn()
{
  // G4cout << "G4MPImanager::WaitBeamOn" << G4endl;
 
  // Extra worker
  if (is_extra_worker_) {
    if ( extra_worker_ ) {
      G4cout << "Calling extra_worker " << G4endl;
      extra_worker_->BeamOn();
    } else {
       G4cout << " !!!! extra_worker_ is not defined " << G4endl;
    }
    return;
  }
 
  G4int buff = 0;
  if ( qbatchmode_ ) {  // valid only in batch mode
    if ( is_master_ ) {
      // receive from each slave
      for (G4int islave = 1; islave < size_; islave++) {
        // G4cout << "calling Irecv for islave " << islave << G4endl; 
        MPI::Request request = COMM_G4COMMAND_.Irecv(&buff, 1, MPI::INT,
                                                     islave, kTAG_G4STATUS);
        while(! request.Test()) {
           ::Wait(1000);
        }
      }
    } else {
      buff = 1;
      // G4cout << "calling send for i " << kRANK_MASTER << G4endl; 
      COMM_G4COMMAND_.Send(&buff, 1, MPI::INT, kRANK_MASTER, kTAG_G4STATUS);
    }
  }
}
 
// --------------------------------------------------------------------------
void G4MPImanager::Print(const G4String& message)
{
  if ( is_master_ ){
    std::cout << message << std::flush;
  } else {
    if ( qfcout_ ) { // output to a file
      fscout_ << message << std::flush;
    } else { // output to stdout
      std::cout << rank_ << ":" << message << std::flush;
    }
  }
}
 
// --------------------------------------------------------------------------
void G4MPImanager::ShowHelp() const
{
  if (is_slave_ ) return;
 
  G4cout << "Geant4 MPI interface" << G4endl;
  G4cout << "usage:" << G4endl;
  G4cout << "<app> [options] [macro file]"
         << G4endl << G4endl;
  G4cout << "   -h, --help              show this message."
         << G4endl;
  G4cout << "   -v, --verbose           show verbose message"
         << G4endl;
  G4cout << "   -i, --init=FNAME        set an init macro file"
         << G4endl;
  G4cout << "   -o, --ofile[=FNAME]     set slave output to a flie"
         << G4endl;
  G4cout << G4endl;
}