Protect concurrent accesses to shared variables in process.

[loba.git] / process.cpp

diff --git a/process.cpp b/process.cpp
index 76524fddaec59b27d08c71c2d492a9383e1f15f9..36bdff07cfbe8d18bf4eaf1b81d5f0cfa6b8f064 100644 (file)
--- a/process.cpp
+++ b/process.cpp
@@ -16,12 +16,15 @@ 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;
 
 int process::process_counter = 0;
 double process::total_load_average;
 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 {
 
 
 namespace {
 


@@ -57,19 +60,21 @@ 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;
     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;
 
     convergence = -1.0;
 

+    proc_mutex->acquire();

     process_counter++;
     process_counter++;

+    total_load_init += real_load;
+    total_load_running += real_load;

     total_load_average = total_load_running / 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;
+    proc_mutex->release();

 
     ctrl_close_pending = data_close_pending = neigh.size();
     close_received = false;
 
     ctrl_close_pending = data_close_pending = neigh.size();
     close_received = false;


@@ -99,7 +104,9 @@ 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)
     xbt_assert(received_load == 0.0,
                "received_load is %g, but should be 0.0 !", received_load);
     if (opt::log_rate < 0)


@@ -117,6 +124,21 @@ process::~process()
     print_loads(true, xbt_log_priority_debug);
 }
 
     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) {


@@ -199,9 +221,12 @@ 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);


@@ -216,23 +241,10 @@ void process::compute_loop()
         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();
-            }
-        }
+        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);


@@ -242,10 +254,13 @@ 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()));


@@ -258,10 +273,34 @@ 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;
+        }
+    } else {
+        if (converged) {
+            XBT_VERB("current load has converged: %g (%.4g%%)",
+                     real_load,  100.0 * load_diff / average);
+            convergence = MSG_get_clock();
+        }
+    }
+}
+

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


@@ -294,7 +333,7 @@ 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.");