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

Private GIT Repository
Display load average and standard deviation at the end.
[loba.git] / process.cpp
index a24d6ebcbfb5eae01f098611a3373a00f06a872f..8b316787b8ae1d8654e0f941337a64c993573f24 100644 (file)
@@ -15,6 +15,7 @@ XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(proc);
 #include "process.h"
 
 double process::total_load_init = 0.0;
+double process::total_load_running = 0.0;
 double process::total_load_exit = 0.0;
 
 process::process(int argc, char* argv[])
@@ -34,6 +35,7 @@ process::process(int argc, char* argv[])
 
     prev_load_broadcast = -1;   // force sending of load on first send()
     expected_load = load;
+    total_load_running += load;
     total_load_init += load;
 
     ctrl_close_pending = data_close_pending = neigh.size();
@@ -44,6 +46,7 @@ process::process(int argc, char* argv[])
     close_received = false;
     may_receive =  (neigh.size() > 0); // the same as (ctrl_close_pending ||
                                        //              data_close_pending)
+    finalizing = false;
     if (may_receive)
         comm.listen();
 
@@ -60,7 +63,7 @@ process::process(int argc, char* argv[])
         oss << neigh.back().get_name();
     }
     LOG1(logp, "Got %s.", oss.str().c_str());
-    print_loads(true, logp);
+    print_loads(false, logp);
 }
 
 process::~process()
@@ -73,52 +76,60 @@ int process::run()
     INFO1("Initial load: %g", load);
     VERB0("Starting...");
     iter = 0;
-    bool one_more = true;
-    do {
-        ++iter;
+    while (true) {
+        if (load > 0.0) {
+            ++iter;
+            if (opt::log_rate && iter % opt::log_rate == 0) {
+                if (opt::bookkeeping)
+                    INFO3("(%u) current load: %g ; expected: %g",
+                          iter, load, expected_load);
+                else
+                    INFO2("(%u) current load: %g",
+                          iter, load);
+            }
 
-        if (opt::log_rate && iter % opt::log_rate == 0) {
             if (opt::bookkeeping)
-                INFO3("(%u) current load: %g ; expected: %g",
-                      iter, load, expected_load);
+                expected_load -= load_balance(expected_load);
             else
-                INFO2("(%u) current load: %g",
-                      iter, load);
-        }
-        print_loads(true, xbt_log_priority_debug);
+                load -= load_balance(load);
 
-        if (opt::bookkeeping)
-            expected_load -= load_balance(expected_load);
-        else
-            load -= load_balance(load);
+            print_loads(true, xbt_log_priority_debug);
 
-        send();
-        compute();
+            send();
+            compute();
 
-// NDS for Need To Send
-#define NDS ((opt::bookkeeping ? expected_load : load) != prev_load_broadcast)
-        do {
-            // General idea: block on receiving unless there is
-            // something to compute, or to send, or we must exit.
+            if (opt::maxiter && iter >= opt::maxiter)
+                break;
+        } else {
+            // send load information, and load when bookkeeping
+            send();
+        }
 
-            // fixme: review this chunk, and remove this NDS macro!
+        // 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));
+        DEBUG1("CALL RECEIVE(%s)", recv_wait? "WAIT": "NO_WAIT");
+        receive(recv_wait? WAIT: NO_WAIT);
 
-            // FIXME: HAD A DEADLOCK HERE...
+        // one of our neighbor is finalizing
+        if (opt::exit_on_close && close_received)
+            break;
 
-            bool recv_wait = (load == 0 && !NDS);
-            DEBUG1("CALL RECEIVE(%s)", recv_wait? "WAIT": "NO_WAIT");
-            receive(recv_wait? WAIT: NO_WAIT);
+        // have no load and cannot receive anything
+        if (load == 0.0 && !may_receive)
+            break;
 
-            if (opt::exit_on_close && close_received)
-                one_more = false;
-            else if (opt::maxiter && iter >= opt::maxiter)
-                one_more = false;
-            
-        } while (one_more && may_receive && load == 0 && !NDS);
-        DEBUG0("RECEIVE LOOP ENDED");
-#undef NDS
+        // fixme: this check should be implemented with a distributed
+        // algorithm, and not a shared global variable!
+        if (100.0 * total_load_running / total_load_init <=
+            opt::load_ratio_threshold) {
+            VERB0("No more load to balance in system, stopping.");
+            break;
+        }
 
-    } while (one_more);
+    }
     VERB0("Going to finalize...");
     finalize();
 
@@ -227,8 +238,8 @@ void process::receive(recv_wait_mode wait)
     //        "NO_WAIT\0WAIT\0\0\0\0WAIT_FOR_CLOSE" + 8 * wait);
     message* msg;
     m_host_t from;
-    bool do_loop = may_receive;
-    while (do_loop && comm.recv(msg, from, wait)) {
+    bool do_wait = (wait != NO_WAIT);
+    while (may_receive && comm.recv(msg, from, do_wait)) {
         switch (msg->get_type()) {
         case message::INFO: {
             neighbor* n = rev_neigh[from];
@@ -238,9 +249,13 @@ void process::receive(recv_wait_mode wait)
         case message::CREDIT:
             expected_load += msg->get_amount();
             break;
-        case message::LOAD:
-            load += msg->get_amount();
+        case message::LOAD: {
+            double ld = msg->get_amount();
+            load += ld;
+            if (finalizing)
+                total_load_running -= ld;
             break;
+        }
         case message::CTRL_CLOSE:
             if (--ctrl_close_pending == 1)
                 comm.next_close_on_ctrl_is_last();
@@ -256,14 +271,14 @@ void process::receive(recv_wait_mode wait)
         }
         delete msg;
         may_receive = (ctrl_close_pending || data_close_pending);
-        do_loop = (wait == WAIT_FOR_CLOSE) && may_receive;
+        do_wait = (wait == WAIT_FOR_CLOSE);
     }
 }
 
 void process::finalize1(neighbor& nb)
 {
     comm.send(nb.get_ctrl_mbox(), new message(message::CTRL_CLOSE, 0.0));
-    comm.send(nb.get_data_mbox(), new message(message::DATA_CLOSE, 0.0));    
+    comm.send(nb.get_data_mbox(), new message(message::DATA_CLOSE, 0.0));
 }
 
 void process::finalize()
@@ -271,6 +286,9 @@ void process::finalize()
     using namespace std::tr1;
     using namespace std::tr1::placeholders;
 
+    finalizing = true;
+    total_load_running -= load;
+
     DEBUG2("send CLOSE to %d neighbor%s.",
            (int )neigh.size(), ESSE(neigh.size()));
     std::for_each(neigh.begin(), neigh.end(),