Neighbor/debt should be protected against concurrent access too.

[loba.git] / process.cpp

diff --git a/process.cpp b/process.cpp
index 8e9f5245fd896a8cd9cfcff05ff51a53b3235c48..21eeb7a194373fdd0b195698dac4360100220f0e 100644 (file)
--- a/process.cpp
+++ b/process.cpp
@@ -21,7 +21,7 @@ double process::total_load_exit = 0.0;
 
 namespace {
 
-    void sleep_until_date(double& date, double duration = 0.0)
+    void sleep_until_date(double& date, double duration)
     {
         double sleep_duration = date - MSG_get_clock();
         if (sleep_duration > 0.0)
@@ -46,8 +46,6 @@ 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_all()
     expected_load = real_load;
     total_load_running += real_load;
@@ -88,7 +86,7 @@ process::~process()
     XBT_INFO("Final load after %d:%d iterations: %g",
              lb_iter, comp_iter, real_load);
     XBT_VERB("Expected load was: %g", expected_load);
-    XBT_VERB("Total computation for this process: %g", comp);
+    XBT_VERB("Total computation for this process: %g", get_comp_amount());
 }
 
 int process::run()
@@ -128,12 +126,17 @@ void process::load_balance_loop()
             ++lb_iter;
         }
 
+        mutex.acquire();
+        if (!opt::bookkeeping)
+            expected_load = real_load - get_sum_of_to_send();
+        // 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_VERB("... expected load: %g", expected_load);
         }
 
-        if (get_load() > 0.0)
+        if (expected_load > 0.0)
             load_balance();
 
         print_loads(true, xbt_log_priority_debug);
@@ -141,7 +144,8 @@ void process::load_balance_loop()
         // send
         std::for_each(neigh.begin(), neigh.end(),
                       bind(&process::ctrl_send, this, _1));
-        prev_load_broadcast = get_load();
+        prev_load_broadcast = expected_load;
+        mutex.release();
 
         sleep_until_date(next_iter_after_date, opt::min_lb_iter_duration);
         ctrl_receive(0.0);
@@ -170,19 +174,23 @@ void process::compute_loop()
     double next_iter_after_date = MSG_get_clock() + opt::min_comp_iter_duration;
     while (still_running()) {
         // receive
+        mutex.acquire();
         if (real_load > 0.0)
             data_receive(0.0);
         else
             data_receive(opt::min_comp_iter_duration);
+        mutex.release();
 
         comm.data_flush(false);
 
-        if (real_load == 0.0)
-            continue;
-
         // send
+        mutex.acquire();
         std::for_each(neigh.begin(), neigh.end(),
                       bind(&process::data_send, this, _1));
+        mutex.release();
+
+        if (real_load == 0.0)
+            continue;
 
         // compute
         ++comp_iter;
@@ -191,7 +199,7 @@ void process::compute_loop()
         TRACE_msg_set_task_category(task, TRACE_CAT_COMP);
         XBT_DEBUG("compute %g flop%s", flops, ESSE(flops));
         MSG_task_execute(task);
-        comp += flops;
+        add_comp_amount(flops);
         MSG_task_destroy(task);
 
         sleep_until_date(next_iter_after_date, opt::min_comp_iter_duration);
@@ -222,6 +230,10 @@ bool process::still_running()
     if (!last_status) {
         /* nop */
 
+    } else if (opt::exit_request) {
+        XBT_VERB("Global exit requested");
+        last_status = false;
+
     } else if (opt::time_limit && MSG_get_clock() >= opt::time_limit) {
         XBT_VERB("Reached time limit: %g/%g", MSG_get_clock(), opt::time_limit);
         last_status = false;
@@ -272,24 +284,27 @@ void process::load_balance()
 
 void process::send(neighbor& nb, double amount)
 {
-    set_load(get_load() - amount);
+    expected_load -= amount;
     nb.set_to_send(nb.get_to_send() + amount);
     nb.set_load(nb.get_load() + amount);
 }
 
 void process::ctrl_send(neighbor& nb)
 {
-    double info_to_send = get_load();
-    if (info_to_send != prev_load_broadcast)
-        comm.ctrl_send(nb.get_ctrl_mbox(),
-                       new message(message::INFO, info_to_send));
+    double info_to_send = expected_load;
+    if (info_to_send != prev_load_broadcast) {
+        message* msg = new message(message::INFO, info_to_send);
+        add_ctrl_send_mesg(msg->get_size());
+        comm.ctrl_send(nb.get_ctrl_mbox(), msg);
+    }
     if (opt::bookkeeping) {
         double debt_to_send = nb.get_to_send();
         if (debt_to_send > 0.0) {
             nb.set_to_send(0.0);
             nb.set_debt(nb.get_debt() + debt_to_send);
-            comm.ctrl_send(nb.get_ctrl_mbox(),
-                           new message(message::CREDIT, debt_to_send));
+            message* msg = new message(message::CREDIT, debt_to_send);
+            add_ctrl_send_mesg(msg->get_size());
+            comm.ctrl_send(nb.get_ctrl_mbox(), msg);
         }
     }
 }
@@ -298,23 +313,33 @@ void process::data_send(neighbor& nb)
 {
     double load_to_send;
     if (opt::bookkeeping) {
-        if (real_load <= nb.get_debt()) {
-            load_to_send = real_load;
+        load_to_send = std::min(real_load, nb.get_debt());
+        if (load_to_send >= opt::min_transfer_amount) {
             nb.set_debt(nb.get_debt() - load_to_send);
-            real_load = 0.0;
-        } else {
-            load_to_send = nb.get_debt();
-            nb.set_debt(0.0);
             real_load -= load_to_send;
+        } else {
+            load_to_send = 0.0;
         }
     } else {
         load_to_send = nb.get_to_send();
-        nb.set_to_send(0.0);
-        // do not update real_load here
+        if (load_to_send >= opt::min_transfer_amount) {
+            nb.set_to_send(0.0);
+            real_load -= load_to_send;
+        } else {
+            load_to_send = 0.0;
+        }
+    }
+    while (load_to_send > 0.0) {
+        double amount;
+        if (opt::max_transfer_amount)
+            amount = std::min(load_to_send, opt::max_transfer_amount);
+        else
+            amount = load_to_send;
+        message* msg = new message(message::LOAD, amount);
+        add_data_send_mesg(msg->get_size());
+        comm.data_send(nb.get_data_mbox(), msg);
+        load_to_send -= amount;
     }
-    if (load_to_send > 0.0)
-        comm.data_send(nb.get_data_mbox(),
-                       new message(message::LOAD, load_to_send));
 }
 
 void process::ctrl_close(neighbor& nb)
@@ -334,6 +359,8 @@ void process::ctrl_receive(double timeout)
 
     XBT_DEBUG("%sblocking receive on ctrl (%g)", "\0non-" + !timeout, timeout);
     while (ctrl_close_pending && comm.ctrl_recv(msg, from, timeout)) {
+        if (msg->get_type() != message::CTRL_CLOSE)
+            add_ctrl_recv_mesg(msg->get_size());
         handle_message(msg, from);
         timeout = 0.0;
     }
@@ -346,6 +373,8 @@ void process::data_receive(double timeout)
 
     XBT_DEBUG("%sblocking receive on data (%g)", "\0non-" + !timeout, timeout);
     while (data_close_pending && comm.data_recv(msg, from, timeout)) {
+        if (msg->get_type() != message::DATA_CLOSE)
+            add_data_recv_mesg(msg->get_size());
         handle_message(msg, from);
         timeout = 0.0;
     }
@@ -356,7 +385,7 @@ void process::handle_message(message* msg, m_host_t from)
     switch (msg->get_type()) {
     case message::INFO: {
         neighbor* n = rev_neigh[from];
-        n->set_load(msg->get_amount());
+        n->set_load(msg->get_amount() + n->get_to_send());
         break;
     }
     case message::CREDIT: