Experimental: use TRY_NO_BACKTRACE if available.

[loba.git] / process.cpp

diff --git a/process.cpp b/process.cpp
index 5d0f98bc5da9fa2532cec0cc0533414a325a146d..01b5e0fcc3ce9caab4dcaefb9b25f0a4a8ad3e6e 100644 (file)
--- a/process.cpp
+++ b/process.cpp
@@ -16,10 +16,19 @@ XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(proc);
 
 #include "process.h"
 
 
 #include "process.h"
 

+mutex_t *process::proc_mutex;
+

 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_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::average_load_ratio;
+double process::load_diff_threshold;
+
+std::atomic<int> process::convergence_counter(0);
+

 namespace {
 
     void sleep_until_date(double& date, double duration)
 namespace {
 
     void sleep_until_date(double& date, double duration)


@@ -56,15 +65,31 @@ process::process(int argc, char* argv[])
 
     prev_load_broadcast = -1;   // force sending of load on first send_all()
     expected_load = real_load;
 
     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;
 
     received_load = 0.0;
 

+    idle_duration = 0.0;
+    convergence = -1.0;
+
+    proc_mutex->acquire();
+    process_counter++;
+    convergence_counter++;
+    total_load_init += real_load;
+    total_load_running += real_load;
+    total_load_average = total_load_running / process_counter;
+    if (opt::avg_load_ratio >= 0.0)
+        average_load_ratio = opt::avg_load_ratio;
+    else
+        average_load_ratio = 100.0 *
+            (process_counter / -opt::avg_load_ratio) / total_load_average;
+    load_diff_threshold = (opt::load_ratio_threshold +
+                           average_load_ratio * total_load_average) / 100.0;
+    proc_mutex->release();
+

     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));


@@ -88,18 +113,41 @@ process::process(int argc, char* argv[])
 process::~process()
 {
     delete lb_thread;
 process::~process()
 {
     delete lb_thread;

+    proc_mutex->acquire();

     total_load_exit += real_load;
     total_load_exit += real_load;

+    proc_mutex->release();

     xbt_assert(received_load == 0.0,
                "received_load is %g, but should be 0.0 !", received_load);
     if (opt::log_rate < 0)
         return;
     xbt_assert(received_load == 0.0,
                "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",
+                 average_load_ratio, convergence);
+    else
+        XBT_INFO("Convergence within %g%% was not achieved",
+                 average_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);
 }
 

+double process::get_iter_deviation() const
+{
+    double average_cost = opt::comp_cost(total_load_average); // fixme: get locked?
+    // 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) {
 int process::run()
 {
     if (opt::log_rate >= 0) {


@@ -144,7 +192,8 @@ 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);
         }
 


@@ -181,9 +230,12 @@ void process::compute_loop()
     using std::placeholders::_1;
 
     double next_iter_after_date = MSG_get_clock() + opt::min_comp_iter_duration;
     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()) {
     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);
 
         // send
         comm.data_flush(false);


@@ -194,10 +246,14 @@ 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;
 

+        convergence_check();
+

         // compute
         // 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);
         ++comp_iter;
         double flops = opt::comp_cost(real_load);
         m_task_t task = MSG_task_create("computation", flops, 0.0, NULL);


@@ -207,10 +263,13 @@ void process::compute_loop()
         add_comp_amount(flops);
         MSG_task_destroy(task);
 
         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__);
         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()));
     finalizing = true;
     XBT_DEBUG("send DATA_CLOSE to %zu neighbor%s",
               neigh.size(), ESSE(neigh.size()));


@@ -223,10 +282,36 @@ void process::compute_loop()
     }
     real_load += received_load;
     received_load = 0.0;
     }
     real_load += received_load;
     received_load = 0.0;

+    proc_mutex->acquire();

     total_load_running -= real_load;
     total_load_running -= real_load;

+    proc_mutex->release();
+    convergence_check();

     comm.data_flush(true);
 }
 
     comm.data_flush(true);
 }
 

+void process::convergence_check()
+{
+    double average = total_load_average; // fixme: get locked?
+    double load_diff = std::fabs(real_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 / average);
+            convergence = -1.0;
+            convergence_counter++;
+        }
+    } else {
+        if (converged) {
+            XBT_VERB("current load has converged: %g (%.4g%%)",
+                     real_load,  100.0 * load_diff / average);
+            convergence = MSG_get_clock();
+            convergence_counter--;
+        }
+    }
+}
+

 bool process::still_running()
 {
     static bool last_status = true;
 bool process::still_running()
 {
     static bool last_status = true;


@@ -250,6 +335,10 @@ bool process::still_running()
         XBT_VERB("Reached comp_maxiter: %d/%d", comp_iter, opt::comp_maxiter);
         last_status = false;
 
         XBT_VERB("Reached comp_maxiter: %d/%d", comp_iter, opt::comp_maxiter);
         last_status = false;
 

+    } else if (opt::exit_on_convergence && convergence_counter == 0) {
+        XBT_VERB("Global convergence detected");
+        last_status = false;
+

     } else if (opt::exit_on_close && close_received) {
         XBT_VERB("Close received");
         last_status = false;
     } else if (opt::exit_on_close && close_received) {
         XBT_VERB("Close received");
         last_status = false;


@@ -259,7 +348,7 @@ bool process::still_running()
         last_status = false;
 
     } else if (100.0 * total_load_running / total_load_init <=
         last_status = false;
 
     } else if (100.0 * total_load_running / total_load_init <=

-               opt::load_ratio_threshold) {
+               opt::load_ratio_threshold) { // fixme: get locked?

         // fixme: this check should be implemented with a distributed
         // algorithm, and not a shared global variable!
         XBT_VERB("No more load to balance in system.");
         // fixme: this check should be implemented with a distributed
         // algorithm, and not a shared global variable!
         XBT_VERB("No more load to balance in system.");