process.cpp

   1 #include <algorithm>
   2 #include <tr1/functional>
   3 #include <iterator>
   4 #include <stdexcept>
   5 #include <sstream>
   6 #include <xbt/log.h>
   7 #include <xbt/time.h>
   8
   9 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(proc);
  10
  11 #include "misc.h"
  12 #include "options.h"
  13 #include "tracing.h"
  14
  15 #include "process.h"
  16
  17 double process::total_load_init = 0.0;
  18 double process::total_load_running = 0.0;
  19 double process::total_load_exit = 0.0;
  20
  21 process::process(int argc, char* argv[])
  22 {
  23     if (argc < 2 || !(std::istringstream(argv[1]) >> real_load))
  24         throw std::invalid_argument("bad or missing initial load parameter");
  25
  26     neigh.assign(argv + 2, argv + argc);
  27
  28     pneigh.reserve(neigh.size());
  29     for (unsigned i = 0 ; i < neigh.size() ; i++) {
  30         neighbor* ptr = &neigh[i];
  31         m_host_t host = MSG_get_host_by_name(ptr->get_name());
  32         pneigh.push_back(ptr);
  33         rev_neigh.insert(std::make_pair(host, ptr));
  34     }
  35
  36     comp = 0.0;
  37
  38     prev_load_broadcast = -1;   // force sending of load on first send_all()
  39     expected_load = real_load;
  40     total_load_running += real_load;
  41     total_load_init += real_load;
  42
  43     ctrl_close_pending = data_close_pending = neigh.size();
  44     close_received = false;
  45     finalizing = false;
  46
  47     comp_iter = lb_iter = 0;
  48
  49     compute_thread = new_msg_thread("compute",
  50                                     std::tr1::bind(&process::compute_loop,
  51                                                    this));
  52
  53     e_xbt_log_priority_t logp = xbt_log_priority_verbose;
  54     if (!LOG_ISENABLED(logp))
  55         return;
  56     std::ostringstream oss;
  57     oss << neigh.size() << " neighbor";
  58     if (!neigh.empty()) {
  59         oss << ESSE(neigh.size()) << ": ";
  60         std::transform(neigh.begin(), neigh.end() - 1,
  61                        std::ostream_iterator<const char*>(oss, ", "),
  62                        std::tr1::mem_fn(&neighbor::get_name));
  63         oss << neigh.back().get_name();
  64     }
  65     XBT_LOG(logp, "Got %s.", oss.str().c_str());
  66     print_loads(false, logp);
  67 }
  68
  69 process::~process()
  70 {
  71     delete compute_thread;
  72     total_load_exit += real_load;
  73     if (opt::bookkeeping) {
  74         XBT_INFO("Final load after %d:%d iterations: %g ; expected: %g",
  75                  lb_iter, comp_iter, real_load, expected_load);
  76     } else {
  77         XBT_INFO("Final load after %d iterations: %g",
  78                  lb_iter, real_load);
  79         if (lb_iter != comp_iter)
  80             XBT_WARN("lb_iter (%d) and comp_iter (%d) differ!",
  81                      lb_iter, comp_iter);
  82     }
  83     XBT_VERB("Total computation for this process: %g", comp);
  84 }
  85
  86 int process::run()
  87 {
  88     XBT_INFO("Initial load: %g", real_load);
  89     XBT_VERB("Starting...");
  90     compute_thread->start();
  91     load_balance_loop();
  92     compute_thread->wait();
  93     XBT_VERB("Done.");
  94     return 0;
  95 }
  96
  97 void process::load_balance_loop()
  98 {
  99     using std::tr1::bind;
 100     using std::tr1::placeholders::_1;
 101
 102     while (still_running()) {
 103         ++lb_iter;
 104
 105         if (opt::log_rate && lb_iter % opt::log_rate == 0) {
 106             if (opt::bookkeeping)
 107                 XBT_INFO("(%u:%u) current load: %g ; expected: %g",
 108                          lb_iter, comp_iter, real_load, expected_load);
 109             else
 110                 XBT_INFO("(%u:%u) current load: %g",
 111                          lb_iter, comp_iter, real_load);
 112         }
 113
 114         if (get_load() > 0.0)
 115             load_balance();
 116
 117         print_loads(true, xbt_log_priority_debug);
 118
 119         // send
 120         std::for_each(neigh.begin(), neigh.end(),
 121                       bind(&process::ctrl_send, this, _1));
 122
 123         // block on receiving unless there is something to compute or
 124         // to send
 125         ctrl_receive(opt::min_lb_iter_duration);
 126
 127         comm.ctrl_flush(false);
 128     }
 129
 130     XBT_VERB("Going to finalize for %s...", __func__);
 131     XBT_DEBUG("send CTRL_CLOSE to %zu neighbor%s",
 132               neigh.size(), ESSE(neigh.size()));
 133     std::for_each(neigh.begin(), neigh.end(),
 134                   bind(&process::ctrl_close, this, _1));
 135     while (ctrl_close_pending) {
 136         comm.ctrl_flush(false);
 137         XBT_DEBUG("waiting for %d CTRL CLOSE", ctrl_close_pending);
 138         ctrl_receive(-1.0);
 139     }
 140     comm.ctrl_flush(true);
 141 }
 142
 143 void process::compute_loop()
 144 {
 145     using std::tr1::bind;
 146     using std::tr1::placeholders::_1;
 147
 148     double next_iter_after_date = 0.0;
 149     while (still_running()) {
 150         // receive
 151         double sleep_duration = real_load
 152             ? std::max(MSG_get_clock() - next_iter_after_date, 0.0)
 153             : opt::min_comp_iter_duration;
 154         data_receive(sleep_duration);
 155
 156         comm.data_flush(false);
 157
 158         if (real_load == 0.0)
 159             continue;
 160
 161         // send
 162         std::for_each(neigh.begin(), neigh.end(),
 163                       bind(&process::data_send, this, _1));
 164
 165         // compute
 166         ++comp_iter;
 167         double flops = opt::comp_cost(real_load);
 168         m_task_t task = MSG_task_create("computation", flops, 0.0, NULL);
 169         TRACE_msg_set_task_category(task, TRACE_CAT_COMP);
 170         XBT_DEBUG("compute %g flop%s", flops, ESSE(flops));
 171         MSG_task_execute(task);
 172         comp += flops;
 173         MSG_task_destroy(task);
 174
 175         next_iter_after_date = MSG_get_clock() + opt::min_comp_iter_duration;
 176     }
 177
 178     XBT_VERB("Going to finalize for %s...", __func__);
 179     XBT_DEBUG("send DATA_CLOSE to %zu neighbor%s",
 180               neigh.size(), ESSE(neigh.size()));
 181     std::for_each(neigh.begin(), neigh.end(),
 182                   bind(&process::data_close, this, _1));
 183     while (data_close_pending) {
 184         comm.data_flush(false);
 185         XBT_DEBUG("waiting for %d DATA CLOSE", data_close_pending);
 186         data_receive(-1.0);
 187     }
 188     comm.data_flush(true);
 189 }
 190
 191 bool process::still_running()
 192 {
 193     static bool last_status = true;
 194
 195     if (!last_status) {
 196         /* nop */
 197
 198     } else if (opt::time_limit && MSG_get_clock() >= opt::time_limit) {
 199         XBT_VERB("Reached time limit: %g/%g", MSG_get_clock(), opt::time_limit);
 200         last_status = false;
 201
 202     } else if (opt::lb_maxiter && lb_iter >= opt::lb_maxiter) {
 203         XBT_VERB("Reached lb_maxiter: %d/%d", lb_iter, opt::lb_maxiter);
 204         last_status = false;
 205
 206     } else if (opt::comp_maxiter && comp_iter >= opt::comp_maxiter) {
 207         XBT_VERB("Reached comp_maxiter: %d/%d", comp_iter, opt::comp_maxiter);
 208         last_status = false;
 209
 210     } else if (opt::exit_on_close && close_received) {
 211         XBT_VERB("Close received");
 212         last_status = false;
 213
 214     } else if (real_load == 0.0 && !data_close_pending) {
 215         XBT_VERB("I'm a poor lonesome process, and I have no load...");
 216         last_status = false;
 217
 218     } else if (100.0 * total_load_running / total_load_init <=
 219                opt::load_ratio_threshold) {
 220         // fixme: this check should be implemented with a distributed
 221         // algorithm, and not a shared global variable!
 222         // fixme: should this chunk be moved before call to receive() ?
 223         XBT_VERB("No more load to balance in system.");
 224         last_status = false;
 225     }
 226
 227     return last_status;
 228 }
 229
 230 void process::load_balance()
 231 {
 232     if (lb_iter == 1)           // warn only once
 233         XBT_WARN("process::load_balance() is a no-op!");
 234 }
 235
 236 void process::send(neighbor& nb, double amount)
 237 {
 238     set_load(get_load() - amount);
 239     nb.set_to_send(nb.get_to_send() + amount);
 240     nb.set_load(nb.get_load() + amount);
 241 }
 242
 243 #define print_loads_generic(vec, verbose, logp, cat)                    \
 244     if (_XBT_LOG_ISENABLEDV((*cat), logp)) {                            \
 245         using std::tr1::bind;                                           \
 246         using std::tr1::placeholders::_1;                               \
 247         XBT_XCLOG(cat, logp, "Neighbor loads:");                        \
 248         std::for_each(vec.begin(), vec.end(),                           \
 249                       bind(&neighbor::print, _1, verbose, logp, cat));  \
 250     } else ((void)0)
 251
 252 void process::print_loads(bool verbose,
 253                           e_xbt_log_priority_t logp,
 254                           xbt_log_category_t cat) const
 255 {
 256     print_loads_generic(neigh, verbose, logp, cat);
 257 }
 258
 259 void process::print_loads_p(bool verbose,
 260                             e_xbt_log_priority_t logp,
 261                             xbt_log_category_t cat) const
 262 {
 263     print_loads_generic(pneigh, verbose, logp, cat);
 264 }
 265
 266 #undef print_loads_generic
 267
 268 void process::ctrl_send(neighbor& nb)
 269 {
 270     double info_to_send = get_load();
 271     if (info_to_send != prev_load_broadcast)
 272         comm.ctrl_send(nb.get_ctrl_mbox(),
 273                        new message(message::INFO, info_to_send));
 274     if (opt::bookkeeping) {
 275         double debt_to_send = nb.get_to_send();
 276         if (debt_to_send > 0.0) {
 277             nb.set_to_send(0.0);
 278             nb.set_debt(nb.get_debt() + debt_to_send);
 279             comm.ctrl_send(nb.get_ctrl_mbox(),
 280                            new message(message::CREDIT, debt_to_send));
 281         }
 282     }
 283 }
 284
 285 void process::data_send(neighbor& nb)
 286 {
 287     double load_to_send;
 288     if (opt::bookkeeping) {
 289         if (real_load <= nb.get_debt()) {
 290             load_to_send = real_load;
 291             nb.set_debt(nb.get_debt() - load_to_send);
 292             real_load = 0.0;
 293         } else {
 294             load_to_send = nb.get_debt();
 295             nb.set_debt(0.0);
 296             real_load -= load_to_send;
 297         }
 298     } else {
 299         load_to_send = nb.get_to_send();
 300         nb.set_to_send(0.0);
 301     }
 302     if (load_to_send > 0.0)
 303         comm.data_send(nb.get_data_mbox(),
 304                        new message(message::LOAD, load_to_send));
 305 }
 306
 307 void process::ctrl_close(neighbor& nb)
 308 {
 309     comm.ctrl_send(nb.get_ctrl_mbox(), new message(message::CTRL_CLOSE, 0.0));
 310 }
 311
 312 void process::data_close(neighbor& nb)
 313 {
 314     comm.data_send(nb.get_data_mbox(), new message(message::DATA_CLOSE, 0.0));
 315 }
 316
 317 void process::ctrl_receive(double timeout)
 318 {
 319     message* msg;
 320     m_host_t from;
 321
 322     XBT_DEBUG("%sblocking receive on ctrl (%g)", "\0non-" + !timeout, timeout);
 323     while (ctrl_close_pending && comm.ctrl_recv(msg, from, timeout)) {
 324         handle_message(msg, from);
 325         timeout = 0.0;
 326     }
 327 }
 328
 329 void process::data_receive(double timeout)
 330 {
 331     message* msg;
 332     m_host_t from;
 333
 334     XBT_DEBUG("%sblocking receive on data (%g)", "\0non-" + !timeout, timeout);
 335     while (data_close_pending && comm.data_recv(msg, from, timeout)) {
 336         handle_message(msg, from);
 337         timeout = 0.0;
 338     }
 339 }
 340
 341 void process::handle_message(message* msg, m_host_t from)
 342 {
 343     switch (msg->get_type()) {
 344     case message::INFO: {
 345         neighbor* n = rev_neigh[from];
 346         n->set_load(msg->get_amount());
 347         break;
 348     }
 349     case message::CREDIT:
 350         expected_load += msg->get_amount();
 351         break;
 352     case message::LOAD: {
 353         double ld = msg->get_amount();
 354         real_load += ld;
 355         if (finalizing)
 356             total_load_running -= ld;
 357         break;
 358     }
 359     case message::CTRL_CLOSE:
 360         ctrl_close_pending--;
 361         close_received = true;
 362         break;
 363     case message::DATA_CLOSE:
 364         data_close_pending--;
 365         close_received = true;
 366         break;
 367     }
 368     delete msg;
 369 }
 370
 371 // Local variables:
 372 // mode: c++
 373 // End: