process.cpp

   1 #include <algorithm>
   2 #include <tr1/functional>
   3 #include <iterator>
   4 #include <numeric>
   5 #include <stdexcept>
   6 #include <sstream>
   7 #include <xbt/log.h>
   8 #include <xbt/time.h>
   9
  10 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(proc);
  11
  12 #include "misc.h"
  13 #include "options.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]) >> 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     prev_load_broadcast = -1;   // force sending of load on first send()
  37     expected_load = load;
  38     total_load_running += load;
  39     total_load_init += load;
  40
  41     ctrl_close_pending = data_close_pending = neigh.size();
  42     if (neigh.size() == 1) {
  43         comm.next_close_on_ctrl_is_last();
  44         comm.next_close_on_data_is_last();
  45     }
  46     close_received = false;
  47     may_receive =  (neigh.size() > 0); // the same as (ctrl_close_pending ||
  48                                        //              data_close_pending)
  49     finalizing = false;
  50     if (may_receive)
  51         comm.listen();
  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     LOG1(logp, "Got %s.", oss.str().c_str());
  66     print_loads(false, logp);
  67 }
  68
  69 process::~process()
  70 {
  71     total_load_exit += load;
  72 }
  73
  74 int process::run()
  75 {
  76     INFO1("Initial load: %g", load);
  77     VERB0("Starting...");
  78     comp_iter = lb_iter = 0;
  79     while (true) {
  80         if (load > 0.0) {
  81             ++comp_iter;
  82             if (opt::log_rate && comp_iter % opt::log_rate == 0) {
  83                 if (opt::bookkeeping)
  84                     INFO4("(%u:%u) current load: %g ; expected: %g",
  85                           comp_iter, lb_iter, load, expected_load);
  86                 else
  87                     INFO2("(%u) current load: %g",
  88                           comp_iter, load);
  89             }
  90
  91             if (opt::bookkeeping)
  92                 expected_load -= load_balance(expected_load);
  93             else
  94                 load -= load_balance(load);
  95
  96             print_loads(true, xbt_log_priority_debug);
  97
  98             send();
  99             compute();
 100
 101         } else {
 102             // send load information, and load when bookkeeping
 103             send();
 104         }
 105
 106         if (opt::comp_maxiter && comp_iter >= opt::comp_maxiter)
 107             break;
 108         if (opt::lb_maxiter && lb_iter >= opt::lb_maxiter)
 109             break;
 110
 111         // block on receiving unless there is something to compute or
 112         // to send
 113         bool recv_wait = (load == 0 &&
 114                           ((opt::bookkeeping ? expected_load : load)
 115                            == prev_load_broadcast));
 116         DEBUG1("CALL RECEIVE(%s)", recv_wait? "WAIT": "NO_WAIT");
 117         receive(recv_wait? WAIT: NO_WAIT);
 118
 119         // one of our neighbor is finalizing
 120         if (opt::exit_on_close && close_received)
 121             break;
 122
 123         // have no load and cannot receive anything
 124         if (load == 0.0 && !may_receive)
 125             break;
 126
 127         // fixme: this check should be implemented with a distributed
 128         // algorithm, and not a shared global variable!
 129         // fixme: should this chunk be moved before call to receive() ?
 130         if (100.0 * total_load_running / total_load_init <=
 131             opt::load_ratio_threshold) {
 132             VERB0("No more load to balance in system, stopping.");
 133             break;
 134         }
 135
 136     }
 137     VERB0("Going to finalize...");
 138     finalize();
 139
 140     /* Open Questions :
 141      * - definition of load on heterogeneous hosts ?
 142      * - how to detect convergence ?
 143      * - how to manage link failures ?
 144      */
 145
 146     VERB0("Done.");
 147     INFO3("Final load after %d iteration%s: %g",
 148           comp_iter, ESSE(comp_iter), load);
 149     if (opt::bookkeeping)
 150         INFO1("Expected load: %g", expected_load);
 151     return 0;
 152 }
 153
 154 double process::sum_of_to_send() const
 155 {
 156     using std::tr1::bind;
 157     using std::tr1::placeholders::_1;
 158     using std::tr1::placeholders::_2;
 159
 160     return std::accumulate(neigh.begin(), neigh.end(), 0.0,
 161                            bind(std::plus<double>(),
 162                                 _1, bind(&neighbor::get_to_send, _2)));
 163 }
 164
 165 double process::load_balance(double /*my_load*/)
 166 {
 167     if (lb_iter == 1)           // warn only once
 168         WARN0("process::load_balance is a no-op!");
 169     return 0.0;
 170 }
 171
 172 void process::compute()
 173 {
 174     if (load > 0.0) {
 175         double duration = opt::comp_cost(load);
 176         m_task_t task = MSG_task_create("computation", duration, 0.0, NULL);
 177         DEBUG2("compute %g flop%s.", duration, ESSE(duration));
 178         MSG_task_execute(task);
 179         MSG_task_destroy(task);
 180     } else {
 181         DEBUG0("nothing to compute !");
 182     }
 183 }
 184
 185 void process::send1_no_bookkeeping(neighbor& nb)
 186 {
 187     if (load != prev_load_broadcast)
 188         comm.send(nb.get_ctrl_mbox(), new message(message::INFO, load));
 189     double load_to_send = nb.get_to_send();
 190     if (load_to_send > 0.0) {
 191         comm.send(nb.get_data_mbox(), new message(message::LOAD, load_to_send));
 192         nb.set_to_send(0.0);
 193     }
 194 }
 195
 196 void process::send1_bookkeeping(neighbor& nb)
 197 {
 198     if (expected_load != prev_load_broadcast)
 199         comm.send(nb.get_ctrl_mbox(),
 200                   new message(message::INFO, expected_load));
 201     double load_to_send;
 202     double new_debt;
 203     double debt_to_send = nb.get_to_send();
 204     if (debt_to_send > 0.0) {
 205         comm.send(nb.get_ctrl_mbox(),
 206                   new message(message::CREDIT, debt_to_send));
 207         nb.set_to_send(0.0);
 208         new_debt = nb.get_debt() + debt_to_send;
 209     } else {
 210         new_debt = nb.get_debt();
 211     }
 212     if (load <= new_debt) {
 213         load_to_send = load;
 214         nb.set_debt(new_debt - load_to_send);
 215         load = 0.0;
 216     } else {
 217         load_to_send = new_debt;
 218         nb.set_debt(0.0);
 219         load -= load_to_send;
 220     }
 221     if (load_to_send > 0.0)
 222         comm.send(nb.get_data_mbox(), new message(message::LOAD, load_to_send));
 223 }
 224
 225 void process::send()
 226 {
 227     using std::tr1::bind;
 228     using std::tr1::placeholders::_1;
 229
 230     if (opt::bookkeeping) {
 231         std::for_each(neigh.begin(), neigh.end(),
 232                       bind(&process::send1_bookkeeping, this, _1));
 233         prev_load_broadcast = expected_load;
 234     } else {
 235         std::for_each(neigh.begin(), neigh.end(),
 236                       bind(&process::send1_no_bookkeeping, this, _1));
 237         prev_load_broadcast = load;
 238     }
 239 }
 240
 241 void process::receive(recv_wait_mode wait)
 242 {
 243     // DEBUG1("go for receive(%s)",
 244     //        "NO_WAIT\0WAIT\0\0\0\0WAIT_FOR_CLOSE" + 8 * wait);
 245     message* msg;
 246     m_host_t from;
 247     bool do_wait = (wait != NO_WAIT);
 248     while (may_receive && comm.recv(msg, from, do_wait)) {
 249         switch (msg->get_type()) {
 250         case message::INFO: {
 251             neighbor* n = rev_neigh[from];
 252             n->set_load(msg->get_amount());
 253             break;
 254         }
 255         case message::CREDIT:
 256             expected_load += msg->get_amount();
 257             break;
 258         case message::LOAD: {
 259             double ld = msg->get_amount();
 260             load += ld;
 261             if (finalizing)
 262                 total_load_running -= ld;
 263             break;
 264         }
 265         case message::CTRL_CLOSE:
 266             if (--ctrl_close_pending == 1)
 267                 comm.next_close_on_ctrl_is_last();
 268             // DEBUG1("ctrl_close_pending = %d", ctrl_close_pending);
 269             close_received = true;
 270             break;
 271         case message::DATA_CLOSE:
 272             if (--data_close_pending == 1)
 273                 comm.next_close_on_data_is_last();
 274             // DEBUG1("data_close_pending = %d", data_close_pending);
 275             close_received = true;
 276             break;
 277         }
 278         delete msg;
 279         may_receive = (ctrl_close_pending || data_close_pending);
 280         do_wait = (wait == WAIT_FOR_CLOSE);
 281     }
 282 }
 283
 284 void process::finalize1(neighbor& nb)
 285 {
 286     comm.send(nb.get_ctrl_mbox(), new message(message::CTRL_CLOSE, 0.0));
 287     comm.send(nb.get_data_mbox(), new message(message::DATA_CLOSE, 0.0));
 288 }
 289
 290 void process::finalize()
 291 {
 292     using std::tr1::bind;
 293     using std::tr1::placeholders::_1;
 294
 295     finalizing = true;
 296     total_load_running -= load;
 297
 298     DEBUG2("send CLOSE to %d neighbor%s.",
 299            (int )neigh.size(), ESSE(neigh.size()));
 300     std::for_each(neigh.begin(), neigh.end(),
 301                   bind(&process::finalize1, this, _1));
 302
 303     DEBUG2("wait for CLOSE from %d neighbor%s.",
 304            (int )neigh.size(), ESSE(neigh.size()));
 305     receive(WAIT_FOR_CLOSE);
 306
 307     comm.flush(true);
 308 }
 309
 310 #define print_loads_generic(vec, verbose, logp, cat)                    \
 311     if (_XBT_LOG_ISENABLEDV((*cat), logp)) {                            \
 312         using std::tr1::bind;                                           \
 313         using std::tr1::placeholders::_1;                               \
 314         XCLOG0(cat, logp, "Neighbor loads:");                           \
 315         std::for_each(vec.begin(), vec.end(),                           \
 316                       bind(&neighbor::print, _1, verbose, logp, cat));  \
 317     } else ((void)0)
 318
 319 void process::print_loads(bool verbose,
 320                           e_xbt_log_priority_t logp,
 321                           xbt_log_category_t cat) const
 322 {
 323     print_loads_generic(neigh, verbose, logp, cat);
 324 }
 325
 326 void process::print_loads_p(bool verbose,
 327                             e_xbt_log_priority_t logp,
 328                             xbt_log_category_t cat) const
 329 {
 330     print_loads_generic(pneigh, verbose, logp, cat);
 331 }
 332
 333 #undef print_loads_generic
 334
 335 // Local variables:
 336 // mode: c++
 337 // End: