]> AND Private Git Repository - loba.git/blobdiff - process.cpp
Logo AND Algorithmique Numérique Distribuée

Private GIT Repository
Merge branch 'master' of ssh://info.iut-bm.univ-fcomte.fr/loba
[loba.git] / process.cpp
index 5f393bdb39abb837b279a21e285ee293483fcbea..57c06612ad84bb9cffaa07123e8c113b0148fad9 100644 (file)
@@ -1,7 +1,6 @@
 #include <algorithm>
 #include <tr1/functional>
 #include <iterator>
-#include <numeric>
 #include <stdexcept>
 #include <sstream>
 #include <xbt/log.h>
@@ -11,6 +10,7 @@ XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(proc);
 
 #include "misc.h"
 #include "options.h"
+#include "tracing.h"
 
 #include "process.h"
 
@@ -20,7 +20,7 @@ double process::total_load_exit = 0.0;
 
 process::process(int argc, char* argv[])
 {
-    if (argc < 2 || !(std::istringstream(argv[1]) >> load))
+    if (argc < 2 || !(std::istringstream(argv[1]) >> real_load))
         throw std::invalid_argument("bad or missing initial load parameter");
 
     neigh.assign(argv + 2, argv + argc);
@@ -33,17 +33,19 @@ process::process(int argc, char* argv[])
         rev_neigh.insert(std::make_pair(host, ptr));
     }
 
-    prev_load_broadcast = -1;   // force sending of load on first send()
-    expected_load = load;
-    total_load_running += load;
-    total_load_init += load;
+    comp = 0.0;
+
+    prev_load_broadcast = -1;   // force sending of load on first send_all()
+    expected_load = real_load;
+    total_load_running += real_load;
+    total_load_init += real_load;
 
     ctrl_close_pending = data_close_pending = neigh.size();
     close_received = false;
-    may_receive =  (neigh.size() > 0); // the same as (ctrl_close_pending ||
-                                       //              data_close_pending)
     finalizing = false;
 
+    comp_iter = lb_iter = 0;
+
     e_xbt_log_priority_t logp = xbt_log_priority_verbose;
     if (!LOG_ISENABLED(logp))
         return;
@@ -62,70 +64,101 @@ process::process(int argc, char* argv[])
 
 process::~process()
 {
-    total_load_exit += load;
+    total_load_exit += real_load;
+    if (opt::bookkeeping) {
+        INFO4("Final load after %d:%d iterations: %g ; expected: %g",
+              lb_iter, comp_iter, real_load, expected_load);
+    } else {
+        INFO2("Final load after %d iterations: %g",
+              lb_iter, real_load);
+        if (lb_iter != comp_iter)
+            WARN2("lb_iter (%d) and comp_iter (%d) differ!",
+                  lb_iter, comp_iter);
+    }
+    VERB1("Total computation for this process: %g", comp);
 }
 
 int process::run()
 {
-    INFO1("Initial load: %g", load);
+    double next_iter_after_date = 0.0;
+    INFO1("Initial load: %g", real_load);
     VERB0("Starting...");
-    comp_iter = lb_iter = 0;
     while (true) {
-        if (load > 0.0) {
-            ++comp_iter;
-            if (opt::log_rate && comp_iter % opt::log_rate == 0) {
+        if (get_load() > 0.0) {
+            double now = MSG_get_clock();
+            if (now < next_iter_after_date)
+                MSG_process_sleep(next_iter_after_date - now);
+            next_iter_after_date = MSG_get_clock() + opt::min_iter_duration;
+
+            ++lb_iter;
+
+            if (opt::log_rate && lb_iter % opt::log_rate == 0) {
                 if (opt::bookkeeping)
                     INFO4("(%u:%u) current load: %g ; expected: %g",
-                          comp_iter, lb_iter, load, expected_load);
+                          lb_iter, comp_iter, real_load, expected_load);
                 else
                     INFO2("(%u) current load: %g",
-                          comp_iter, load);
+                          lb_iter, real_load);
             }
 
-            if (opt::bookkeeping)
-                expected_load -= load_balance(expected_load);
-            else
-                load -= load_balance(load);
+            load_balance();
 
             print_loads(true, xbt_log_priority_debug);
+        }
 
-            send();
+        // send load information, and load (data) if any
+        send_all();
+        if (real_load > 0.0) {
+            ++comp_iter;
             compute();
-
-        } else {
-            // send load information, and load when bookkeeping
-            send();
         }
 
-        if (opt::comp_maxiter && comp_iter >= opt::comp_maxiter)
+        if (opt::lb_maxiter && lb_iter >= opt::lb_maxiter) {
+            VERB2("Reached lb_maxiter: %d/%d", lb_iter, opt::lb_maxiter);
+            break;
+        }
+        if (opt::comp_maxiter && comp_iter >= opt::comp_maxiter) {
+            VERB2("Reached comp_maxiter: %d/%d", comp_iter, opt::comp_maxiter);
             break;
-        if (opt::lb_maxiter && lb_iter >= opt::lb_maxiter)
+        }
+        if (opt::time_limit && MSG_get_clock() >= opt::time_limit) {
+            VERB2("Reached time limit: %g/%g", MSG_get_clock(), opt::time_limit);
             break;
+        }
 
         // block on receiving unless there is something to compute or
         // to send
-        bool recv_wait = (load == 0 &&
-                          ((opt::bookkeeping ? expected_load : load)
-                           == prev_load_broadcast));
-        receive(recv_wait? WAIT: NO_WAIT);
+        double timeout;
+        if (real_load != 0 || get_load() != prev_load_broadcast)
+            timeout = 0.0;
+        else if (opt::min_iter_duration)
+            timeout = opt::min_iter_duration;
+        else
+            timeout = 1.0;
+        receive(timeout);
 
         // one of our neighbor is finalizing
-        if (opt::exit_on_close && close_received)
+        if (opt::exit_on_close && close_received) {
+            VERB0("Close received");
             break;
+        }
 
         // have no load and cannot receive anything
-        if (load == 0.0 && !may_receive)
+        if (real_load == 0.0 && !may_receive()) {
+            VERB0("I'm a poor lonesome process, and I have no load...");
             break;
+        }
 
         // fixme: this check should be implemented with a distributed
         // algorithm, and not a shared global variable!
         // fixme: should this chunk be moved before call to receive() ?
         if (100.0 * total_load_running / total_load_init <=
             opt::load_ratio_threshold) {
-            VERB0("No more load to balance in system, stopping.");
+            VERB0("No more load to balance in system.");
             break;
+        } else {
+            DEBUG1("still %g load to balance, continuing...", total_load_running);
         }
-
     }
     VERB0("Going to finalize...");
     finalize();
@@ -137,48 +170,41 @@ int process::run()
      */
 
     VERB0("Done.");
-    INFO3("Final load after %d iteration%s: %g",
-          comp_iter, ESSE(comp_iter), load);
-    if (opt::bookkeeping)
-        INFO1("Expected load: %g", expected_load);
     return 0;
 }
 
-double process::sum_of_to_send() const
-{
-    using std::tr1::bind;
-    using std::tr1::placeholders::_1;
-    using std::tr1::placeholders::_2;
-
-    return std::accumulate(neigh.begin(), neigh.end(), 0.0,
-                           bind(std::plus<double>(),
-                                _1, bind(&neighbor::get_to_send, _2)));
-}
-
-double process::load_balance(double /*my_load*/)
+void process::load_balance()
 {
     if (lb_iter == 1)           // warn only once
-        WARN0("process::load_balance is a no-op!");
-    return 0.0;
+        WARN0("process::load_balance() is a no-op!");
 }
 
 void process::compute()
 {
-    if (load > 0.0) {
-        double duration = opt::comp_cost(load);
-        m_task_t task = MSG_task_create("computation", duration, 0.0, NULL);
-        DEBUG2("compute %g flop%s", duration, ESSE(duration));
+    if (real_load > 0.0) {
+        double flops = opt::comp_cost(real_load);
+        m_task_t task = MSG_task_create("computation", flops, 0.0, NULL);
+        TRACE_msg_set_task_category(task, TRACE_CAT_COMP);
+        DEBUG2("compute %g flop%s", flops, ESSE(flops));
         MSG_task_execute(task);
+        comp += flops;
         MSG_task_destroy(task);
     } else {
         DEBUG0("nothing to compute !");
     }
 }
 
+void process::send(neighbor& nb, double amount)
+{
+    set_load(get_load() - amount);
+    nb.set_to_send(nb.get_to_send() + amount);
+    nb.set_load(nb.get_load() + amount); // fixme: make this optional?
+}
+
 void process::send1_no_bookkeeping(neighbor& nb)
 {
-    if (load != prev_load_broadcast)
-        comm.send(nb.get_ctrl_mbox(), new message(message::INFO, load));
+    if (real_load != prev_load_broadcast)
+        comm.send(nb.get_ctrl_mbox(), new message(message::INFO, real_load));
     double load_to_send = nb.get_to_send();
     if (load_to_send > 0.0) {
         comm.send(nb.get_data_mbox(), new message(message::LOAD, load_to_send));
@@ -202,20 +228,20 @@ void process::send1_bookkeeping(neighbor& nb)
     } else {
         new_debt = nb.get_debt();
     }
-    if (load <= new_debt) {
-        load_to_send = load;
+    if (real_load <= new_debt) {
+        load_to_send = real_load;
         nb.set_debt(new_debt - load_to_send);
-        load = 0.0;
+        real_load = 0.0;
     } else {
         load_to_send = new_debt;
         nb.set_debt(0.0);
-        load -= load_to_send;
+        real_load -= load_to_send;
     }
     if (load_to_send > 0.0)
         comm.send(nb.get_data_mbox(), new message(message::LOAD, load_to_send));
 }
 
-void process::send()
+void process::send_all()
 {
     using std::tr1::bind;
     using std::tr1::placeholders::_1;
@@ -227,20 +253,18 @@ void process::send()
     } else {
         std::for_each(neigh.begin(), neigh.end(),
                       bind(&process::send1_no_bookkeeping, this, _1));
-        prev_load_broadcast = load;
+        prev_load_broadcast = real_load;
     }
+    comm.flush(false);
 }
 
-void process::receive(recv_wait_mode wait)
+void process::receive(double timeout)
 {
     message* msg;
     m_host_t from;
 
-    DEBUG1("go for receive(%s)",
-           "NO_WAIT\0WAIT\0\0\0\0WAIT_FOR_CLOSE" + 8 * wait);
-
-    bool do_wait = (wait != NO_WAIT);
-    while (may_receive && comm.recv(msg, from, do_wait)) {
+    DEBUG2("%sblocking receive (%g)", "\0non-" + !timeout, timeout);
+    while (may_receive() && comm.recv(msg, from, timeout)) {
         switch (msg->get_type()) {
         case message::INFO: {
             neighbor* n = rev_neigh[from];
@@ -252,26 +276,24 @@ void process::receive(recv_wait_mode wait)
             break;
         case message::LOAD: {
             double ld = msg->get_amount();
-            load += ld;
+            real_load += ld;
             if (finalizing)
                 total_load_running -= ld;
             break;
         }
         case message::CTRL_CLOSE:
             ctrl_close_pending--;
-            DEBUG1("ctrl_close_pending = %d", ctrl_close_pending);
             close_received = true;
             break;
         case message::DATA_CLOSE:
             data_close_pending--;
-            DEBUG1("data_close_pending = %d", data_close_pending);
             close_received = true;
             break;
         }
         delete msg;
-        may_receive = (ctrl_close_pending || data_close_pending);
-        do_wait = (wait == WAIT_FOR_CLOSE);
+        timeout = 0.0;          // only wait on first recv
     }
+    comm.flush(false);
 }
 
 void process::finalize1(neighbor& nb)
@@ -286,16 +308,19 @@ void process::finalize()
     using std::tr1::placeholders::_1;
 
     finalizing = true;
-    total_load_running -= load;
+    total_load_running -= real_load;
 
     DEBUG2("send CLOSE to %lu neighbor%s",
            (unsigned long )neigh.size(), ESSE(neigh.size()));
     std::for_each(neigh.begin(), neigh.end(),
                   bind(&process::finalize1, this, _1));
 
-    DEBUG2("wait for CLOSE from %lu neighbor%s",
-           (unsigned long )neigh.size(), ESSE(neigh.size()));
-    receive(WAIT_FOR_CLOSE);
+    while (may_receive()) {
+        comm.flush(false);
+        DEBUG2("waiting for %d CTRL and %d DATA CLOSE",
+               ctrl_close_pending, data_close_pending);
+        receive(-1.0);
+    }
 
     comm.flush(true);
 }