source: gs2-extensions/parallel-building/trunk/src/src/mpiimport-src/mpiimport.cpp@ 24699

// Master-worker program to read in a list of files and invoke
// import on each separately using manifest files in Greenstone 3,
// with synchronisation using OpenMPI
//
// Hussein Suleman
// 1 July 2010

// 2011MAR - added a bunch of commands to aid (my) understanding.
//         - made site argument optional to support GS2.
//         - moving manifest writing code into the rank=0 thread. This will
//           remove the artificial limit on epoc size caused by size of
//           message buffer between controller and child threads. [jmt12]
// 2011SEP - changing where the manifest files are written so I have the
//           ability to set up the location as a RAMDisk. [jmt12]

#include "mpi.h"

#include <stdio.h>
#include <stdlib.h>

#include <fstream>
#include <iostream>
#include <sstream>
#include <string>
#include <vector>

using namespace std;

#define BUFFERSIZE 512

int
main ( int argc, char *argv [] )
{
  int numtasks, rank, rc;            // MPI variables

  if (5 != argc && argc != 6 )
  {
    cerr << "Usage: " << argv[0] << " filelist epoch gsdlhome collection [site]" << endl;
    exit(-1);
  }

  char *filelist = argv[1];          // list of filenames
  char *epochStr = argv[2];          // number of files per task
  int epoch = atoi (epochStr);
  char *gsdlhomedir = argv[3];      // location of import script
  char *collection = argv[4];        // Greenstone collection
  char *site = "";
  if (argc == 6)
  {
    site = argv[5];              // Greenstone site
  }

  // start MPI environment
  rc = MPI_Init(&argc,&argv);
  if (rc != MPI_SUCCESS)
  {
    printf ("Error starting MPI program. Terminating.\n");
    MPI_Abort(MPI_COMM_WORLD, rc);
  }

  // We'll handle errors ourselves
  MPI_Errhandler_set(MPI_COMM_WORLD, MPI_ERRORS_RETURN);

  // get MPI variables: number of processors and processor number
  MPI_Status stat;
  MPI_Comm_size(MPI_COMM_WORLD, &numtasks);
  MPI_Comm_rank(MPI_COMM_WORLD, &rank);

  // master node processing
  if (rank == 0)
  {
    char incoming[BUFFERSIZE];          // buffer for acknowledgments
    char buffer[BUFFERSIZE];         // buffer to send tasks
    MPI_Request request[BUFFERSIZE];    // request monitor for all tasks
    MPI_Status status[BUFFERSIZE];      // status monitor for all tasks
    int actualTasks = 0;         // number of processors running

    vector<char *> manifest_files; // Keep track of temp manifest filenames

    // open file listing filenames to process
    ifstream infile;
    infile.open (filelist);
    string line;

    // set initial status of all processors to idle
    for ( int j=0; j<BUFFERSIZE; j++ )
    {
      incoming[j] = ' ';
    }

    // scan through contents of file listing
    int manifest_file_count = 1;
    while (!infile.eof ())
    {
      // get a filename
      getline (infile, line);
      if (line.length() > 0)
      {
        // search for idle processor
        int dest=0;
        int found = 0;
        while ((dest<(numtasks-1)) && (found == 0))
        {
          if (incoming[dest] == ' ')
          {
            found = 1;
          }
          else
          {
            dest++;
          }
        }

        // if no idle processor, wait for one to become idle
        if (found == 0)
        {
          MPI_Waitany (numtasks-1, request, &dest, status);
        }

        // construct manifest filename
    stringstream manifestfilename_strstr;
    manifestfilename_strstr << gsdlhomedir << "/collect/" << collection << "/tmp/manifest." << manifest_file_count << ".xml";
    string manifestfilename_str = manifestfilename_strstr.str();
    char *manifestfilename = new char [manifestfilename_str.size() + 1];
    strcpy(manifestfilename, manifestfilename_str.c_str());
        ofstream manifestfile;
        if (manifestfilename != NULL)
        {
          // create manifest file
          manifestfile.open (manifestfilename);
          if (manifestfile.fail())
          {
            cerr << "Fatal Error! Failed to open temporary manifest file for writing: " << manifestfile << endl;
          }
        }
        else
        {
          cerr << "Fatal Error! Failed to open temporary manifest file for writing: " << manifestfile << endl;
        }
        manifestfile << "<Manifest><Index>" << endl;

        // add the first filename to the instruction
        manifestfile << "<Filename>" << line << "</Filename>" << endl;
        int epochCounter = epoch;

        // if epoch>1 and more filenames, add more filenames
        while ((epochCounter > 1) && (!infile.eof ()))
        {
          getline (infile, line);
          if (line.length () > 0)
          {
            manifestfile << "<Filename>" << line << "</Filename>" << endl;
            epochCounter--;
          }
        }

        manifestfile << "</Index></Manifest>" << endl;
        manifestfile.close ();

        // Store manifest filepath so we can remove it later
        manifest_files.push_back(manifestfilename);

        // Send the manifest filename as the instruction
        sprintf(buffer, "%s", manifestfilename);

        // mark processors as busy
        incoming[dest] = 'B';
        // send out the job to the processor
        MPI_Send (&buffer, strlen (buffer)+1, MPI_CHAR, dest+1, 1, MPI_COMM_WORLD);
        // wait for a done acknowledgement
        MPI_Irecv (&incoming[dest], 1, MPI_CHAR, dest+1, 1, MPI_COMM_WORLD, &request[dest]);
        // update counter of actual tasks
        if (dest > actualTasks)
        {
          actualTasks = dest;
        }
    // increase number of manifest files processed
    manifest_file_count++;
      }
    }

    infile.close();

    // wait until all outstanding tasks are completed
    int dest;
    for ( int k=0; k<actualTasks; k++ )
    {
      MPI_Waitany (actualTasks, request, &dest, status);
    }

    // send message to end all processing engines
    char endstr[5] = "end";
    for ( int i=1; i<numtasks; i++ )
    {
      MPI_Send (endstr, 4, MPI_CHAR, i, 1, MPI_COMM_WORLD);
    }

    // Free up manifest files
    for ( int i = 0; i < manifest_files.size(); i++)
    {
      free(manifest_files[i]);
    }

  }
  // slave node processing
  else
  {
    char incoming[BUFFERSIZE];

    do
    {
      // wait for instruction from master
      int resval = MPI_Recv (&incoming, BUFFERSIZE, MPI_CHAR, 0, 1, MPI_COMM_WORLD, &stat);
      if (resval != MPI_SUCCESS)
      {
        printf ("Error when recieving message from master [%d]... Terminating.\n", resval);
        MPI_Abort(MPI_COMM_WORLD, rc);
      }
      if (strcmp (incoming, "end") != 0)
      {
        // process a received job
        cout << "Processing [" << rank << ":" << incoming << "]" << endl;

        // create Greenstone import command
        // - incoming now contains the name of the manifest file to process
        char command[2048];
        if (site != "")
        {
          sprintf (command, "%s/bin/script/import.pl -keepold -manifest %s -site %s %s", gsdlhomedir, incoming, site, collection);
        }
        else
        {
          sprintf (command, "%s/bin/script/import.pl -keepold -manifest %s %s", gsdlhomedir, incoming, collection);
        }
        cout << "**** cmd = " << command << endl;

        // invoke Greenstone import with manifest file
        system (command);

        char line = ' ';
        // send completed message
        MPI_Send (&line, 1, MPI_CHAR, 0, 1, MPI_COMM_WORLD);
      }
    }
    while (strcmp (incoming, "end") != 0);
    // stop when "end" instruction is received
  }

  // clean up MPI environment
  MPI_Finalize();
}