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

Private GIT Repository
Cosmetics: use a litteral constant of type double.
[loba.git] / process.cpp
index 5d0f98bc5da9fa2532cec0cc0533414a325a146d..76524fddaec59b27d08c71c2d492a9383e1f15f9 100644 (file)
@@ -20,6 +20,9 @@ double process::total_load_init = 0.0;
 double process::total_load_running = 0.0;
 double process::total_load_exit = 0.0;
 
 double process::total_load_running = 0.0;
 double process::total_load_exit = 0.0;
 
+int process::process_counter = 0;
+double process::total_load_average;
+
 namespace {
 
     void sleep_until_date(double& date, double duration)
 namespace {
 
     void sleep_until_date(double& date, double duration)
@@ -54,17 +57,25 @@ process::process(int argc, char* argv[])
         rev_neigh.insert(std::make_pair(host, ptr));
     }
 
         rev_neigh.insert(std::make_pair(host, ptr));
     }
 
+    // Note: there should not be race condition with the current
+    // version of Simgrid, when updating the global variables.
+
     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;
     received_load = 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;
     received_load = 0.0;
 
+    convergence = -1.0;
+
+    process_counter++;
+    total_load_average = total_load_running / process_counter;
+
     ctrl_close_pending = data_close_pending = neigh.size();
     close_received = false;
     finalizing = false;
 
     ctrl_close_pending = data_close_pending = neigh.size();
     close_received = false;
     finalizing = false;
 
-    comp_iter = lb_iter = 0;
+    all_comp_iter = comp_iter = lb_iter = 0;
 
     lb_thread = new_msg_thread("loba",
                                std::bind(&process::load_balance_loop, this));
 
     lb_thread = new_msg_thread("loba",
                                std::bind(&process::load_balance_loop, this));
@@ -93,8 +104,14 @@ process::~process()
                "received_load is %g, but should be 0.0 !", received_load);
     if (opt::log_rate < 0)
         return;
                "received_load is %g, but should be 0.0 !", received_load);
     if (opt::log_rate < 0)
         return;
-    XBT_INFO("Final load after %d:%d iterations: %g",
-             lb_iter, comp_iter, real_load);
+    XBT_INFO("Final load after %d:%d:%d iterations: %g",
+             lb_iter, comp_iter, all_comp_iter, real_load);
+    if (convergence >= 0.0)
+        XBT_INFO("Convergence within %g%% was achieved at time %g",
+                 opt::avg_load_ratio, convergence);
+    else
+        XBT_INFO("Convergence within %g%% was not achieved",
+                 opt::avg_load_ratio);
     XBT_VERB("Expected load was: %g", expected_load);
     XBT_VERB("Total computation for this process: %g", get_comp_amount());
     print_loads(true, xbt_log_priority_debug);
     XBT_VERB("Expected load was: %g", expected_load);
     XBT_VERB("Total computation for this process: %g", get_comp_amount());
     print_loads(true, xbt_log_priority_debug);
@@ -144,7 +161,8 @@ void process::load_balance_loop()
         // nothing to do with opt::bookkeeping
 
         if (opt::log_rate && lb_iter % opt::log_rate == 0) {
         // nothing to do with opt::bookkeeping
 
         if (opt::log_rate && lb_iter % opt::log_rate == 0) {
-            XBT_INFO("(%u:%u) current load: %g", lb_iter, comp_iter, real_load);
+            XBT_INFO("(%u:%u:%u) current load: %g",
+                     lb_iter, comp_iter, all_comp_iter, real_load);
             XBT_VERB("... expected load: %g", expected_load);
         }
 
             XBT_VERB("... expected load: %g", expected_load);
         }
 
@@ -194,9 +212,26 @@ void process::compute_loop()
                       std::bind(&process::data_send, this, _1));
         mutex.release();
 
                       std::bind(&process::data_send, this, _1));
         mutex.release();
 
+        ++all_comp_iter;
         if (real_load == 0.0)
             continue;
 
         if (real_load == 0.0)
             continue;
 
+        double load_ratio =
+            100.0 * std::fabs(real_load / total_load_average - 1.0);
+        if (convergence >= 0.0) {
+            if (load_ratio > opt::avg_load_ratio) {
+                XBT_VERB("current load has diverged: %g (%.4g%%)",
+                         real_load, load_ratio);
+                convergence = -1.0;
+            }
+        } else {
+            if (load_ratio <= opt::avg_load_ratio) {
+                XBT_VERB("current load has converged: %g (%.4g%%)",
+                         real_load, load_ratio);
+                convergence = MSG_get_clock();
+            }
+        }
+
         // compute
         ++comp_iter;
         double flops = opt::comp_cost(real_load);
         // compute
         ++comp_iter;
         double flops = opt::comp_cost(real_load);