Document a bug with parallel executions.

[loba.git] / process.cpp

diff --git a/process.cpp b/process.cpp
index 07372dc0515f763e8f16729d4437c6b94c55df5d..c85b6051a6a4224452f22ddce086740f310ed61e 100644 (file)
--- a/process.cpp
+++ b/process.cpp
@@ -20,6 +20,10 @@ double process::total_load_init = 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;
+double process::load_diff_threshold;
+
 namespace {
 
     void sleep_until_date(double& date, double duration)
@@ -54,12 +58,23 @@ process::process(int argc, char* argv[])
         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;
 
+    idle_duration = 0.0;
+    convergence = -1.0;
+
+    process_counter++;
+    total_load_average = total_load_running / process_counter;
+    load_diff_threshold = (opt::load_ratio_threshold +
+                           opt::avg_load_ratio * total_load_average) / 100.0;
+
     ctrl_close_pending = data_close_pending = neigh.size();
     close_received = false;
     finalizing = false;
@@ -95,11 +110,32 @@ process::~process()
         return;
     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);
 }
 
+double process::get_iter_deviation() const
+{
+    double average_cost = opt::comp_cost(total_load_average);
+    // Do not count idle periods
+    double comp_iter_opt = acc.comp_amount / average_cost;
+/*
+    // Add iterations that could have been achieved while beeing idle
+    // (kept for documentation)
+    double self_speed = MSG_get_host_speed(MSG_host_self());
+    double average_duration = average_cost / self_speed;
+    comp_iter_opt += idle_duration / average_duration;
+*/
+    return comp_iter - comp_iter_opt;
+}
+
 int process::run()
 {
     if (opt::log_rate >= 0) {
@@ -182,9 +218,12 @@ void process::compute_loop()
     using std::placeholders::_1;
 
     double next_iter_after_date = MSG_get_clock() + opt::min_comp_iter_duration;
+    double idle_since_date = 0.0;
     while (still_running()) {
-        // receive (do not block if there is something to compute)
-        data_receive(real_load > 0.0 ? 0.0 : opt::min_comp_iter_duration);
+        // receive
+        // if there is something to compute, do not block
+        // else, block the duration of an *lb* iteration
+        data_receive(real_load > 0.0 ? 0.0 : opt::min_lb_iter_duration);
 
         // send
         comm.data_flush(false);
@@ -199,7 +238,10 @@ void process::compute_loop()
         if (real_load == 0.0)
             continue;
 
+        convergence_check();
+
         // compute
+        idle_duration += MSG_get_clock() - idle_since_date;
         ++comp_iter;
         double flops = opt::comp_cost(real_load);
         m_task_t task = MSG_task_create("computation", flops, 0.0, NULL);
@@ -209,10 +251,13 @@ void process::compute_loop()
         add_comp_amount(flops);
         MSG_task_destroy(task);
 
+        idle_since_date = MSG_get_clock();
+
         sleep_until_date(next_iter_after_date, opt::min_comp_iter_duration);
     }
 
     XBT_VERB("Going to finalize for %s...", __func__);
+    // Note: idle duration is not counted during finalization
     finalizing = true;
     XBT_DEBUG("send DATA_CLOSE to %zu neighbor%s",
               neigh.size(), ESSE(neigh.size()));
@@ -226,9 +271,30 @@ void process::compute_loop()
     real_load += received_load;
     received_load = 0.0;
     total_load_running -= real_load;
+    convergence_check();
     comm.data_flush(true);
 }
 
+void process::convergence_check()
+{
+    double load_diff = std::fabs(real_load - total_load_average);
+    bool converged = load_diff <= load_diff_threshold;
+
+    if (convergence >= 0.0) {
+        if (!converged) {
+            XBT_VERB("current load has diverged: %g (%.4g%%)",
+                     real_load, 100.0 * load_diff / total_load_average);
+            convergence = -1.0;
+        }
+    } else {
+        if (converged) {
+            XBT_VERB("current load has converged: %g (%.4g%%)",
+                     real_load,  100.0 * load_diff / total_load_average);
+            convergence = MSG_get_clock();
+        }
+    }
+}
+
 bool process::still_running()
 {
     static bool last_status = true;