static void send_mrsg_data (msg_task_t msg)
{
char mailbox[MAILBOX_ALIAS_SIZE];
double data_size;
size_t my_id;
mrsg_task_info_t ti;
my_id = get_mrsg_worker_id (MSG_host_self ());
sprintf (mailbox, TASK_MRSG_MAILBOX,
get_mrsg_worker_id (MSG_task_get_source (msg)),
MSG_process_get_PID (MSG_task_get_sender (msg)));
if (mrsg_message_is (msg, SMS_GET_MRSG_CHUNK))
{
MSG_task_dsend (MSG_task_create ("DATA-C", 0.0, config_mrsg.mrsg_chunk_size, NULL), mailbox, NULL);
}
else if (mrsg_message_is (msg, SMS_GET_INTER_MRSG_PAIRS))
{
ti = (mrsg_task_info_t) MSG_task_get_data (msg);
data_size = job_mrsg.map_output[my_id][ti->mrsg_tid] - ti->map_output_copied[my_id];
MSG_task_dsend (MSG_task_create ("DATA-IP", 0.0, data_size, NULL), mailbox, NULL);
}
MSG_task_destroy (msg);
}
/** @brief Main master function. */
int master (int argc, char* argv[])
{
heartbeat_t heartbeat;
msg_error_t status;
msg_host_t worker;
msg_task_t msg = NULL;
size_t wid;
task_info_t ti;
double total_cpu_time = 0.0;
double total_task_time = 0.0;
print_config ();
XBT_INFO ("JOB BEGIN"); XBT_INFO (" ");
tasks_log = fopen ("tasks.csv", "w");
fprintf (tasks_log, "task_id,phase,worker_id,time,action,shuffle_end\n");
while (job.tasks_pending[MAP] + job.tasks_pending[REDUCE] > 0)
{
msg = NULL;
status = receive (&msg, MASTER_MAILBOX);
if (status == MSG_OK)
{
worker = MSG_task_get_source (msg);
wid = get_worker_id (worker);
if (message_is (msg, SMS_HEARTBEAT))
{
heartbeat = &job.heartbeats[wid];
if (is_straggler (worker))
{
set_speculative_tasks (worker);
}
else
{
if (heartbeat->slots_av[MAP] > 0)
send_scheduler_task(MAP, wid);
if (heartbeat->slots_av[REDUCE] > 0)
send_scheduler_task(REDUCE, wid);
}
}
else if (message_is (msg, SMS_TASK_DONE))
{
ti = (task_info_t) MSG_task_get_data (msg);
if (job.task_status[ti->phase][ti->id] != T_STATUS_DONE)
{
job.task_status[ti->phase][ti->id] = T_STATUS_DONE;
finish_all_task_copies (ti);
job.tasks_pending[ti->phase]--;
if (job.tasks_pending[ti->phase] <= 0)
{
XBT_INFO (" ");
XBT_INFO ("%s PHASE DONE", (ti->phase==MAP?"MAP":"REDUCE"));
XBT_INFO (" ");
}
ti->finished_time = MSG_get_clock();
ti->elapsed_time = ti->finished_time - ti->start_time;
total_task_time += ti->elapsed_time;
total_cpu_time += ti->cpu_time;
}
xbt_free_ref (&ti);
}
MSG_task_destroy (msg);
}
}
fclose (tasks_log);
job.finished = 1;
print_config ();
print_stats ();
XBT_INFO ("JOB END");
XBT_INFO ("\tclock_time: %f", MSG_get_clock());
XBT_INFO ("\ttotal_task_time: %f(%f)", total_task_time, total_task_time / MSG_get_clock());
XBT_INFO ("\ttotal_cpu_time: %f(%f)", total_cpu_time, total_cpu_time / MSG_get_clock());
return 0;
}
/* Master Process */
int master(int argc, char *argv[])
{
char * key;
struct HdmsgHost *hdmsg_host;
xbt_dict_cursor_t cursor = NULL;
int i = 0;
long remaining_inits = 0;
long remaining_mappers = 0;
long remaining_shufflers = 0;
long remaining_reducers = 0;
long expected_messages = 0;
msg_comm_t res_irecv;
msg_task_t task_com;
msg_task_t *tasks = xbt_new(msg_task_t, number_of_workers);
xbt_dynar_t comms = xbt_dynar_new(sizeof(msg_comm_t), NULL);
XBT_INFO("INITIALIZATION BEGIN");
// Initialize processes (mappers, shufflers, and reducers) on each host
xbt_dict_foreach(hosts, cursor, key, hdmsg_host)
{
if (hdmsg_host->is_worker)
{
MSG_process_create("Init", initializeProcs, NULL, hdmsg_host->host);
tasks[remaining_inits] = NULL;
res_irecv = MSG_task_irecv(&tasks[remaining_inits], "master");
xbt_dynar_push_as(comms, msg_comm_t, res_irecv);
remaining_inits++;
}
}
while (!xbt_dynar_is_empty(comms))
{
xbt_dynar_remove_at(comms, MSG_comm_waitany(comms), &res_irecv);
task_com = MSG_comm_get_task(res_irecv);
if (!strcmp(MSG_task_get_name(task_com), "init_exit"))
{
msg_host_t h = MSG_task_get_source(task_com);
MSG_task_destroy(task_com);
const char *host_name = MSG_host_get_name(h);
struct HdmsgHost *hdmsg_host = xbt_dict_get(hosts, host_name);
remaining_mappers += get_mapper_count(hdmsg_host);
remaining_shufflers += get_shuffler_count(hdmsg_host);
remaining_reducers += get_reducer_count(hdmsg_host);
remaining_inits--;
if (remaining_inits == 0)
{
XBT_INFO("INITIALIZATION COMPLETE");
// Add an extra message to account for the message sent when the shuffle phase begins
expected_messages = 1 + remaining_mappers + remaining_shufflers + remaining_reducers;
free(tasks);
tasks = xbt_new(msg_task_t, expected_messages);
for (i = 0; i < expected_messages; i++)
{
tasks[i] = NULL;
res_irecv = MSG_task_irecv(&tasks[i], "master");
xbt_dynar_push_as(comms, msg_comm_t, res_irecv);
}
XBT_INFO("MAP PHASE BEGIN");
// Activate Mappers
xbt_dict_foreach(hosts, cursor, key, hdmsg_host)
{
activate_mappers(hdmsg_host);
}
}
