Bug fixed: use a timeout on receive.

[loba.git] / process.cpp

diff --git a/process.cpp b/process.cpp
index 9f46f8f8d0c1f067cf34b725747771c063b8740c..0f535763510940caae9cf9373b36a83b4c1f0694 100644 (file)
--- a/process.cpp
+++ b/process.cpp
@@ -33,6 +33,8 @@ process::process(int argc, char* argv[])
         rev_neigh.insert(std::make_pair(host, ptr));
     }
 
+    comp = 0.0;
+
     prev_load_broadcast = -1;   // force sending of load on first send()
     expected_load = load;
     total_load_running += load;
@@ -75,7 +77,7 @@ int process::run()
             if (now < next_iter_after_date){
                 double delay = next_iter_after_date - now;
                 DEBUG1("sleeping for %g s", delay);
-                MSG_process_sleep(next_iter_after_date - now);
+                MSG_process_sleep(delay);
             }
             next_iter_after_date = MSG_get_clock() + opt::min_iter_duration;
         }
@@ -121,8 +123,14 @@ int process::run()
 
         // block on receiving unless there is something to compute or
         // to send
-        bool wait = (load == 0 && lb_load() == prev_load_broadcast);
-        receive(wait);
+        double timeout;
+        if (load != 0 || lb_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) {
@@ -143,6 +151,8 @@ int process::run()
             opt::load_ratio_threshold) {
             VERB0("No more load to balance in system.");
             break;
+        } else {
+            DEBUG1("still %g load to balance, continuing...", total_load_running);
         }
     }
     VERB0("Going to finalize...");
@@ -189,10 +199,11 @@ double process::load_balance(double /*my_load*/)
 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));
+        double flops = opt::comp_cost(load);
+        m_task_t task = MSG_task_create("computation", flops, 0.0, NULL);
+        DEBUG2("compute %g flop%s", flops, ESSE(flops));
         MSG_task_execute(task);
+        comp += flops;
         MSG_task_destroy(task);
     } else {
         DEBUG0("nothing to compute !");
@@ -256,12 +267,13 @@ void process::send()
     comm.flush(false);
 }
 
-void process::receive(bool wait)
+void process::receive(double timeout)
 {
     message* msg;
     m_host_t from;
 
-    while (may_receive() && comm.recv(msg, from, 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];
@@ -288,7 +300,7 @@ void process::receive(bool wait)
             break;
         }
         delete msg;
-        wait = false;           // only wait on first recv
+        timeout = 0.0;          // only wait on first recv
     }
     comm.flush(false);
 }
@@ -316,7 +328,7 @@ void process::finalize()
            (unsigned long )neigh.size(), ESSE(neigh.size()));
     while (may_receive()) {
         comm.flush(false);
-        receive(true);
+        receive(-1.0);
     }
 
     comm.flush(true);