source: gs2-extensions/parallel-building/trunk/src/src/mpibuildcol-src/mpibuildcol.cpp@ 24624

// Uses OpenMPI to allow several Greenstone build passes to be run in parallel.
// Author: John Thompson, 2010NOV18
// Adapted from code by: Hussein Suleman, 2010JUL01

// 0. Initialization
// Uncomment to display debug information
#define DEBUG 1
// Comment out to disable XML reading
#define TINYXML 1
// Fixed buffer size - needs to be fixed size to pass between processes
#define BUFFERSIZE 256
// Special message tags
#define EXITTAG     0
#define SUCCESSTAG  1

#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <iostream>
#include <queue>
#include <sstream>
#include <stack>
#include <string>
#include <vector>

#include <mpi.h>

#ifdef TINYXML
#include "tinyxml.h"
#endif

using namespace std;

struct task_t
{
  int id;
  int prerequisite; /* -1 for no prerequisite */
  string command;
};
// Lets call an order group of tasks a recipe, shall we?
typedef vector<task_t> recipe_t;

/* Function Prototypes */
static void debugPrint(int,string);
static void masterProcess(const char*);
static void parseRecipe(const char*, recipe_t*);
static void workerProcess(int);
#ifdef TINYXML
static int recurseRecipeXML(recipe_t*, TiXmlNode*, int, int);
#endif
/**
 */
int
main(int argc, char **argv)
{
  if (argc != 2)
  {
    cerr << "usage: mpibuild.cpp <recipe xml>" << endl;
    return 0;
  }
  // 1. Initialize MPI
  MPI_Init(&argc, &argv);
  // 2. Find out my identity in the 'world'
  int myrank;
  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
  if (myrank == 0)
  {
    debugPrint(-1,"===== MPIBuild  =====");
  }
  // - and dispatch to the appropriate chunk of code
  if (myrank == 0)
  {
    masterProcess(argv[1]);
  }
  else
  {
    workerProcess(myrank);
  }
  // Shut down MPI
  MPI_Finalize();
  if (myrank == 0)
  {
    debugPrint(-1,"===== Complete! =====");
  }
  return 0;
}
/** main(int argc, char **argv) **/

/**
 */
static void
debugPrint(int source, string message)
{
#ifdef DEBUG
  time_t seconds = time (NULL);
  cout << "[" << seconds << "]";
  if (source == 0)
  {
    cout << "[Master] ";
  }
  else if (source > 0)
  {
    cout << "[Worker" << source << "] ";
  }
  cout << message << endl;
#endif
}
/** debugPrint(int,string) **/

/** The activity undertaken by the master 'thread'
 */
static void
masterProcess(const char* recipe_xml_path)
{
  debugPrint(0, "starting");
  // 0. Initialize
  debugPrint(0, "initializing");
  // - find out how many processes there are in the default communicator
  int number_of_processes;
  MPI_Comm_size(MPI_COMM_WORLD, &number_of_processes);
  // - sanity check
  if (number_of_processes < 2)
  {
    cerr << "Error! Minimum number of processes is 2" << endl;
  }
  // - remember that '0' is the master processor
  int number_of_workers = number_of_processes - 1;
  // - initialize the pool of idle worker processes...
  stack<int> idle_workers;
  for (int i = 1; i <= number_of_workers; i++)
  {
    idle_workers.push(i);
  }
  // - we also need a queue of tasks ready to be undertaken. We use a queue as
  //   the recipe should be if preferred order of execution - and so might be
  //   slightly more efficient if run in that order
  queue<task_t> ready_tasks;

  // 1. Parse in the 'recipe' that controls what configurations of Greenstone
  //    build can be called and when. We end up with a 'queue' of tasks, each
  //    with a unique id (int), a prequisite (int) if required, and a shell
  //    command (string).
  debugPrint(0, "parsing recipe");
  recipe_t recipe;
  parseRecipe(recipe_xml_path, &recipe);
  // - go ahead and define some iterators
  recipe_t::iterator recipe_begin_itr = recipe.begin();
  recipe_t::iterator recipe_end_itr = recipe.end();
  recipe_t::iterator recipe_current_itr;

  // - now look through recipe for tasks that are ready to go immediate (i.e.
  //   have no prerequisite tasks) and add them to the queue of ready tasks
  debugPrint(0, "queueing ready tasks");
  recipe_current_itr = recipe_begin_itr;
  while (recipe_current_itr != recipe_end_itr)
  {
    task_t a_task = *recipe_current_itr;
    if (a_task.prerequisite == -1)
    {
      ready_tasks.push(a_task);
    }
    ++recipe_current_itr;
  }

  // 2. We have a 'pool' of idle workers and a list of ready tasks. Start by
  //    iterating through the tasks assigning them to workers until we have
  //    either run out of actionable tasks (although some tasks may become
  //    actionable once others finish) or have exhausted the pool of workers.
  debugPrint(0, "initial task assignment");
  while (ready_tasks.size() > 0 && idle_workers.size() > 0)
  {
    task_t a_task = ready_tasks.front();
    ready_tasks.pop();
    // - grab an idle worker
    int worker_id = idle_workers.top();
    idle_workers.pop();
    // - create a fixed size buffer to store the command string in
    char buffer[BUFFERSIZE];
    sprintf(buffer, "%s", a_task.command.c_str());
    // - seed the slaves; send one unit of work to each slave.
    stringstream strstr;
    strstr << " - assigning task " << a_task.id << " to worker " << worker_id;
    debugPrint(0, strstr.str());
    MPI_Send(&buffer,           // message buffer containing command
             strlen(buffer)+1,  // command string length
             MPI_CHAR,          // data item is a character array
             worker_id,         // destination process rank
             a_task.id,         // we use the task id as the worktag!
             MPI_COMM_WORLD);   // default communicator
  }
  // - by now we have either a) assigned all the ready jobs to workers
  //   or, b) we've run out of idle workers.
  stringstream str1;
  str1 << ready_tasks.size();
  debugPrint(0, str1.str() + " ready tasks remaining");
  stringstream str2;
  str2 << idle_workers.size();
  debugPrint(0, str2.str() + " idle workers remaining");

  // 3. Assuming we have at least one worker that is busy (so the not all the
  //    workers are sitting in the idle pool), we wait/block until we receive
  //    feedback from a worker process (feedback that includes the identifier
  //    of the task just completed). We go through the recipe and queue any
  //    dependant tasks as ready to go, while returning the worker to the idle
  //    pool. We then essentially repeat the process of matching task to
  //    working trying to exhaust either the queue of ready tasks or of will
  //    slaves. (note: maybe optimize this later)
  while (idle_workers.size() < number_of_workers)
  {
    // - wait until a worker completes and replies
    debugPrint(0, "waiting until some worker process responds");
    int task_id;
    MPI_Status status;
    MPI_Recv(&task_id,          // the identifier of the task completed
             1,                 // a single interger expected
             MPI_INT,           // data item is an int
             MPI_ANY_SOURCE,    // receive from any sender
             MPI_ANY_TAG,       // any type of message tag
             MPI_COMM_WORLD,    // default communicator
             &status);          // info about the received message
    // - loop through the tasks, looking for any that were waiting for
    //   this task to be completed, and add them to the ready queue.
    debugPrint(0, "queuing any tasks that are now ready");
    recipe_current_itr = recipe_begin_itr;
    while (recipe_current_itr < recipe_end_itr)
    {
      task_t a_task = *recipe_current_itr;
      if (a_task.prerequisite == task_id)
      {
        ready_tasks.push(a_task);
      }
      ++recipe_current_itr;
    }
    // - status contains the identifier of the worker process...
    int worker_id = status.MPI_SOURCE;
    // - ...which we add to the pool of idle workers
    idle_workers.push(worker_id);
    // - now we try, once again, to match tasks to workers until we run out of
    //   one or the other
    debugPrint(0, "task assignment");
    while (ready_tasks.size() > 0 && idle_workers.size() > 0)
    {
      task_t a_task = ready_tasks.front();
      ready_tasks.pop();
      // - grab an idle worker
      int worker_id = idle_workers.top();
      idle_workers.pop();
      // - create a fixed size buffer to store the command string in
      char buffer[BUFFERSIZE];
      if (sprintf(buffer, "%s", a_task.command.c_str()) < 0)
      {
        cerr << "Error! Failed to write command string into transport buffer." << endl;
      }
      // - send the task to the worker
      stringstream strstr;
      strstr << " - assigning task " << a_task.id << " to worker " << worker_id;
      debugPrint(0, strstr.str());
      MPI_Send(&buffer,           // message buffer containing command
               strlen(buffer)+1,  // command string length
               MPI_CHAR,          // data item is a character array
               worker_id,         // destination process rank
               a_task.id,         // we use the task id as the worktag!
               MPI_COMM_WORLD);   // default communicator
    }
    stringstream str3;
    str3 << ready_tasks.size();
    debugPrint(0, str3.str() + " ready tasks remaining");
    stringstream str4;
    str4 << idle_workers.size();
    debugPrint(0, str4.str() + " idle workers remaining");
  }
  // - we can do some sanity checking here. For instance, there should be no
  //   tasks left in the ready queue at this point and all workers should be
  //   sitting in the idle pool
  if (ready_tasks.size() > 0)
  {
    cerr << "Error! Processing supposedly complete but tasks are still pending!" << endl;
  }
  if (idle_workers.size() != number_of_workers)
  {
    cerr << "Error! Processing supposedly complete but workers are still busy!" << endl;
  }

  // 4. By now all workers have returned, and should we waiting idle in the
  //    pool. Iterate over all workers telling them to exit by using the
  //    special EXITTAG.
  for (int worker_id = 1; worker_id <= number_of_workers; worker_id++)
  {
    MPI_Send(0,               // we don't intend to do any processing
             0,               // zero data items
             MPI_INT,         // data item is an integer
             worker_id,       // destination process identifier
             EXITTAG,         // message flag indicating workers must exit
             MPI_COMM_WORLD); // default communicator
  }

  // 5. At this point all the slaves have expired and the master process is
  //    complete
  debugPrint(0, "exiting");
}
/** masterProcess(const char *) **/

/** Reads in an XML file, and builds an (ordered) list of tasks which I've
 *  termed a recipe.
 *  - note: when parsing, code should verify that no command exceeds BUFFERSIZE
 *          characters in length
 */
static void
parseRecipe(const char* recipe_xml_path, recipe_t* recipe)
{
#ifdef TINYXML
  TiXmlDocument recipe_xml_doc(recipe_xml_path);
  bool is_loaded = recipe_xml_doc.LoadFile();
  if (is_loaded)
  {
    // - top node is always the document (in this case <Recipe>) so we loop
    //   through it's child elements (which should be the no-prequisite tasks)
    TiXmlElement *root = recipe_xml_doc.RootElement();
    int task_count = 0;
    TiXmlNode* element;
    for ( element = root->FirstChild(); element != 0; element = element->NextSibling())
    {
      task_count++;
      task_count = recurseRecipeXML(recipe, element, -1, task_count);
    }
  }
  else
  {
    cerr << "Error! Failed to open/parse XML file: " << recipe_xml_path << endl;
  }

#else
  // - dummy data for a start. We'll emulate a (complex) MG build with four
  //   indexes
  task_t task1;
  task1.id = 1;
  task1.prerequisite = -1;
  task1.command = "buildcol.pl -keepold -verbosity 3 -mode compress_text tdb0005000 >> build.log 2>&1";
  recipe->push_back(task1);
  task_t task2;
  task2.id = 2;
  task2.prerequisite = -1;
  task2.command = "buildcol.pl -keepold -verbosity 3 -mode infodb tdb0005000 >> build.log 2>&1";
  recipe->push_back(task2);
  task_t task3;
  task3.id = 3;
  task3.prerequisite = 1;
  task3.command = "buildcol.pl -keepold -verbosity 3 -mode build_index -indexname document:text tdb0005000 >> build.log 2>&1";
  recipe->push_back(task3);
  task_t task4;
  task4.id = 4;
  task4.prerequisite = 1;
  task4.command = "buildcol.pl -keepold -verbosity 3 -mode build_index -indexname document:Title tdb0005000 >> build.log 2>&1";
  recipe->push_back(task4);
  task_t task5;
  task5.id = 5;
  task5.prerequisite = 1;
  task5.command = "buildcol.pl -keepold -verbosity 3 -mode build_index -indexname document:Source tdb0005000 >> build.log 2>&1";
  recipe->push_back(task5);
  task_t task6;
  task6.id = 6;
  task6.prerequisite = 1;
  task6.command = "buildcol.pl -keepold -verbosity 3 -mode build_index -indexname document:Title,Source tdb0005000 >> build.log 2>&1";
  recipe->push_back(task6);
  task_t task7;
  task7.id = 7;
  task7.prerequisite = 1;
  task7.command = "buildcol.pl -keepold -verbosity 3 -mode build_index -indexname document:text,Title,Source tdb0005000 >> build.log 2>&1";
  recipe->push_back(task7);
#endif

#ifdef DEBUG
  debugPrint(0, " * The Recipe!");
  // If we are debugging, lets print out the recipe
  recipe_t::iterator recipe_begin_itr = (*recipe).begin();
  recipe_t::iterator recipe_end_itr = (*recipe).end();
  recipe_t::iterator recipe_current_itr;
  for(recipe_current_itr = recipe_begin_itr; recipe_current_itr < recipe_end_itr; ++recipe_current_itr)
  {
    task_t a_task = (task_t)(*recipe_current_itr);
    stringstream recipestrstr;
    recipestrstr << " Task " << a_task.id << ": " << a_task.command << " [Prequisite: " << a_task.prerequisite << "]";
    debugPrint(0, recipestrstr.str());
  }
#endif
}
/** parseRecipe(const char*, recipe_t*) **/

/**
 */
#ifdef TINYXML
static int recurseRecipeXML(recipe_t* recipe, TiXmlNode* element, int parent_id, int task_count)
{
  // A. Loop through this task elements children. At least one subelement
  //    should be the command, while all other subelements should be
  //    tasks that need to be further recursed.
  TiXmlNode* child;
  int task_id = task_count;
  bool found_command = false; // We can only have one command per task id
  for (child = element->FirstChild(); child != 0; child = child->NextSibling())
  {
    // - we are only interested in element nodes
    if (child->Type() == TiXmlNode::TINYXML_ELEMENT)
    {
      stringstream element_name_stream;
      element_name_stream << child->Value();
      string element_name = element_name_stream.str();
      if (element_name == "Command" && !found_command)
      {
        // - retrieve the text child node
        for ( TiXmlNode *text_grandchild = child->FirstChild(); text_grandchild != 0; text_grandchild = text_grandchild->NextSibling())
    {
          if (text_grandchild->Type() == TiXmlNode::TINYXML_TEXT)
          {
            TiXmlText* text_command = text_grandchild->ToText();
            stringstream commandstrstr;
            commandstrstr << text_command->Value() << " 2>&1";
            // - and create the new task given the command string and
            //   accounting for any preqrequisites via the the parent task id
            //   (begin non-zero)
            task_t a_task;
            a_task.id = task_id;
            a_task.prerequisite = parent_id;
            a_task.command = commandstrstr.str();
            recipe->push_back(a_task);
            // - and prevent any further commands being associated with this
            //   task id
            found_command = true;
          }
    }
      }
      else if (element_name == "Task")
      {
        task_count++;
        recurseRecipeXML(recipe, child, task_id, task_count);
      }
    }
  }
  return task_count;
}
#endif
/** recurseRecipeXML(recipe_t*, TiXmlNode*, int, int) **/

/** Each worker is responsible for executing a particular command in a shell,
 *  waiting until the command is complete, and then relaying to the master that
 *  work is complete.
 *  @param worker_id  An integer containing the unique identifier (rank) of the
 *                    worker process.
 *  @return void
 */
static void
workerProcess(int worker_id)
{
  debugPrint(worker_id, "starting");
  // 0. Worker loops tirelessly until asked to exit (note that the mpi_recv
  //    command below is blocking - so no tight-loop thrashing)
  while (1)
  {
    // 1. Receive a message from the master process
    char command[BUFFERSIZE];
    MPI_Status status;
    MPI_Recv(&command,       // buffer in which to store the command string
             BUFFERSIZE,     // we expect (at most) BUFFERSIZE characters
             MPI_CHAR,       // we expect a char array
             0,              // recieve from the master process only
             MPI_ANY_TAG,    // accept any incoming tag
             MPI_COMM_WORLD, // default communicator
             &status);
    // - watch for the special exit tag, and end the worker process if detected
    if (status.MPI_TAG == EXITTAG)
    {
      debugPrint(worker_id, "exiting");
      return;
    }
    // - otherwise the tag actually tells us the task id - which we need to
    //   reply back with later
    int task_id = status.MPI_TAG;
    // - action the command in the buffer
    string message = "processing: ";
    message += command;
    debugPrint(worker_id, message);
    // asynchronous
    //system(command);
    FILE *pipe = popen(command, "r");
    if (!pipe)
    {
      cerr << "Warning! Failed to open pipe to command: " << command << endl;
    }
    else
    {
      char buffer[1024];
      while(fgets(buffer, sizeof(buffer), pipe) != NULL)
      {
        cout << buffer;
      }
      pclose(pipe);
    }
    // 2. Send a reply back containing the id of the task just completed
    debugPrint(worker_id, "complete");
    MPI_Send(&task_id,
             1,
             MPI_INT,
             0,
             SUCCESSTAG,
             MPI_COMM_WORLD);
  }
}
/** workerProcess(int) **/



/*
#include "mpi.h"

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

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

using namespace std;

#define KILOBUF 512
#define MEGABUF 10240

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

  if (argc != 3 )
  {
    cerr << "Usage: " << argv[0] << " gsdlhome collection" << endl;
    exit(-1);
  }

  // Greenstone home directory
  char *gsdl_home_dir = argv[1];
  // Short collection name
  char *collection = argv[2];
                
  // start MPI environment
  int mpi_status = MPI_Init(&argc, &argv);
  if (mpi_status != MPI_SUCCESS)
  {
    printf ("Error starting MPI program. Terminating.\n");
    MPI_Abort(MPI_COMM_WORLD, mpi_status);
  }
  
  // get MPI variables: number of processors and processor number
  MPI_Status stat;
  MPI_Comm_size(MPI_COMM_WORLD, &num_tasks);
  MPI_Comm_rank(MPI_COMM_WORLD, &rank);

  // The Master node (dispatcher) has a rank of zero. All child processes have
  // higher ranks.
  if (rank == 0)
  {
    cout << "===== MPI Build =====" << endl;
    cout << "Uses OpenMPI to allow several Greenstone build passes to be run in parallel." << endl;
    cout << endl;
    cout << "Number of Processors: " << num_tasks << endl;
    cout << "Rank of current process: " << rank << endl;
    cout << endl;

    cout << "[0] Dispatcher" << endl;
    // buffer for acknowledgments
    // - Is there some reason this buffer isn't actually set to be numtasks in
    //   size?
    char incoming[KILOBUF];
    // Buffer to send tasks
    char buffer[KILOBUF];
    // Request monitor for all tasks
    MPI_Request request[KILOBUF];
    // Status monitor for all tasks
    MPI_Status status[KILOBUF];
    // Number of processors running
    int actual_tasks = 0;

    // Set initial status of all processors to idle
    cout << " - initializing child tasks to idle" << endl;
    for ( int i = 0;  i < KILOBUF; i++ )
    {
      incoming[i] = ' ';
    }

    // In the future this would be where the process reads in settings from the
    // collect.cfg to determine how many passes are needed and thus what tasks
    // there are available for children to process. For the moment I'll just
    // hardcode the short three task list that essentially invokes each of the
    // three standard Greenstone build modes - compress_text, build_indexes and
    // make_infodb.
    cout << " - populating task list... ";
    // List of pending tasks to undertake
    vector<string> tasks;
    tasks.push_back("infodb");
    tasks.push_back("compress_text");
    tasks.push_back("build_index");
    cout << "found " << tasks.size() << " tasks" << endl;

    // For each pending task
    for(int j = 0; j < tasks.size(); j++)
    {
      // Search for idle processor
      cout << " - searching for idle processor" << endl;
      int dest=0;
      int found = 0;
      while ( (dest < ( num_tasks - 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(num_tasks-1, request, &dest, status);
      }

      // Write the tasks mode flag as the instruction to the child
      sprintf(buffer, "%s", tasks[j].c_str());

      // Mark processors as busy
      incoming[dest] = 'B';

      cout << " - asking child process to execute: " << tasks[j] << endl;

      // Send out the job to the processor
      MPI_Send(&buffer, strlen (buffer)+1, MPI_CHAR, dest+1, 1, MPI_COMM_WORLD);

      // I'm guess there is some kind of 'fork' either here or above. Somehow
      // this command 'waits' for a response from the child thread. In the
      // usual case the response is a single space (' ') that overwrites the
      // 'B' in the incoming string at the position that matches the child
      // thread number. (Presumably you could count the B's in the incoming
      // string to determine running threads... but we don't).
      MPI_Irecv(&incoming[dest], 1, MPI_CHAR, dest+1, 1, MPI_COMM_WORLD, &request[dest]);
      cout << "[DEBUG] When does this code actually execute? [cmd: " << buffer << "]" << endl;

      // Update counter of actual tasks
      if (dest > actual_tasks)
      {
        actual_tasks = dest;
      }
    }

    // Wait until all outstanding tasks are completed
    cout << " - waiting for outstanding tasks" << endl;
    int dest;
    for ( int k = 0; k < actual_tasks; k++ )
    {
      MPI_Waitany(actual_tasks, request, &dest, status);
    }

    // Send message to end all processing engines
    cout << " - ask all child processes to terminate" << endl;
    char endstr[5] = "end";
    for ( int l = 1; l < num_tasks; l++ )
    {
      MPI_Send(endstr, 4, MPI_CHAR, l, 1, MPI_COMM_WORLD);
    }
  }
  // Slave node processing
  else
  {
    cout << "[" << rank << "] Child Process" << endl;
    char incoming[KILOBUF];

    do
    {   
      // wait for instruction from master
      MPI_Recv(&incoming, KILOBUF, MPI_CHAR, 0, 1, MPI_COMM_WORLD, &stat);
      if (strcmp(incoming, "end") != 0)
      {
        // Process a received job
        cout << " + processing command: " << incoming << endl;
          
        // Create Greenstone import command
        char command[2048];
        sprintf(command, "%s/bin/script/buildcol.pl -keepold -verbosity 3 -mode %s %s", gsdl_home_dir, incoming, collection);
        cout << " + Greenstone buildcol command: " << 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);
        cout << " + done [cmd: " << command << "]" << endl;
      }
    }
    // stop when "end" instruction is received
    while (strcmp(incoming, "end") != 0);

    cout << " + child process terminating" << endl;
  }

  // clean up MPI environment
  MPI_Finalize();
  if (0 == rank)
  {
    cout << "Complete!" << endl << endl;
  }
}
*/