Deadlock fix, and other changes.

[loba.git] / process.cpp

diff --git a/process.cpp b/process.cpp
index 1a9066d66d013ba234f097fcffa5661cbba2a476..8b316787b8ae1d8654e0f941337a64c993573f24 100644 (file)
--- a/process.cpp
+++ b/process.cpp
@@ -1,36 +1,53 @@
-#include "process.h"
-

 #include <algorithm>
 #include <tr1/functional>
 #include <iterator>
 #include <algorithm>
 #include <tr1/functional>
 #include <iterator>

+#include <numeric>

 #include <stdexcept>
 #include <sstream>
 #include <xbt/log.h>
 #include <xbt/time.h>
 #include <stdexcept>
 #include <sstream>
 #include <xbt/log.h>
 #include <xbt/time.h>

+
+XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(proc);
+

 #include "misc.h"
 #include "options.h"
 
 #include "misc.h"
 #include "options.h"
 

-XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simu);
+#include "process.h"
+
+double process::total_load_init = 0.0;
+double process::total_load_running = 0.0;
+double process::total_load_exit = 0.0;

 
 process::process(int argc, char* argv[])
 {
     if (argc < 2 || !(std::istringstream(argv[1]) >> load))
 
 process::process(int argc, char* argv[])
 {
     if (argc < 2 || !(std::istringstream(argv[1]) >> load))

-        throw std::invalid_argument("bad or missing initial load");
+        throw std::invalid_argument("bad or missing initial load parameter");

 
     neigh.assign(argv + 2, argv + argc);
 
 
     neigh.assign(argv + 2, argv + argc);
 

-    std::for_each(neigh.begin(), neigh.end(),
-                  std::tr1::bind(&process::insert_neighbor_in_map,
-                                 this, std::tr1::placeholders::_1));
+    pneigh.reserve(neigh.size());
+    for (unsigned i = 0 ; i < neigh.size() ; i++) {
+        neighbor* ptr = &neigh[i];
+        m_host_t host = MSG_get_host_by_name(ptr->get_name());
+        pneigh.push_back(ptr);
+        rev_neigh.insert(std::make_pair(host, ptr));
+    }

 
 

+    prev_load_broadcast = -1;   // force sending of load on first send()

     expected_load = load;
     expected_load = load;

+    total_load_running += load;
+    total_load_init += load;

 
     ctrl_close_pending = data_close_pending = neigh.size();
     if (neigh.size() == 1) {
         comm.next_close_on_ctrl_is_last();
         comm.next_close_on_data_is_last();
     }
 
     ctrl_close_pending = data_close_pending = neigh.size();
     if (neigh.size() == 1) {
         comm.next_close_on_ctrl_is_last();
         comm.next_close_on_data_is_last();
     }

-    if (neigh.size() > 0)
+    close_received = false;
+    may_receive =  (neigh.size() > 0); // the same as (ctrl_close_pending ||
+                                       //              data_close_pending)
+    finalizing = false;
+    if (may_receive)

         comm.listen();
 
     e_xbt_log_priority_t logp = xbt_log_priority_verbose;
         comm.listen();
 
     e_xbt_log_priority_t logp = xbt_log_priority_verbose;


@@ -46,44 +63,72 @@ process::process(int argc, char* argv[])
         oss << neigh.back().get_name();
     }
     LOG1(logp, "Got %s.", oss.str().c_str());
         oss << neigh.back().get_name();
     }
     LOG1(logp, "Got %s.", oss.str().c_str());

-    print_loads(logp);
+    print_loads(false, logp);

 }
 
 }
 

-int process::run()
+process::~process()

 {
 {

-    bool one_more = true;
+    total_load_exit += load;
+}

 
 

+int process::run()
+{

     INFO1("Initial load: %g", load);
     VERB0("Starting...");
     iter = 0;
     INFO1("Initial load: %g", load);
     VERB0("Starting...");
     iter = 0;

-    while (one_more) {
-        ++iter;
+    while (true) {
+        if (load > 0.0) {
+            ++iter;
+            if (opt::log_rate && iter % opt::log_rate == 0) {
+                if (opt::bookkeeping)
+                    INFO3("(%u) current load: %g ; expected: %g",
+                          iter, load, expected_load);
+                else
+                    INFO2("(%u) current load: %g",
+                          iter, load);
+            }

 
 

-        if (opt::log_rate && iter % opt::log_rate == 0) {

             if (opt::bookkeeping)
             if (opt::bookkeeping)

-                INFO3("(%u) current load: %g ; expected: %g",
-                      iter, load, expected_load);
+                expected_load -= load_balance(expected_load);

             else
             else

-                INFO2("(%u) current load: %g",
-                      iter, load);
-        }
+                load -= load_balance(load);

 
 

-        compute();
+            print_loads(true, xbt_log_priority_debug);

 
 

-        bool close_received = !receive(false);
+            send();
+            compute();

 
 

-        if (opt::bookkeeping)
-            expected_load -= load_balance(expected_load);
-        else
-            load -= load_balance(load);
+            if (opt::maxiter && iter >= opt::maxiter)
+                break;
+        } else {
+            // send load information, and load when bookkeeping
+            send();
+        }

 
 

-        send();
-        comm.flush(false);
+        // block on receiving unless there is something to compute or
+        // to send
+        bool recv_wait = (load == 0 &&
+                          ((opt::bookkeeping ? expected_load : load)
+                           == prev_load_broadcast));
+        DEBUG1("CALL RECEIVE(%s)", recv_wait? "WAIT": "NO_WAIT");
+        receive(recv_wait? WAIT: NO_WAIT);

 
 

+        // one of our neighbor is finalizing

         if (opt::exit_on_close && close_received)
         if (opt::exit_on_close && close_received)

-            one_more = false;
-        if (opt::maxiter && iter >= opt::maxiter)
-            one_more = false;
+            break;
+
+        // have no load and cannot receive anything
+        if (load == 0.0 && !may_receive)
+            break;
+
+        // fixme: this check should be implemented with a distributed
+        // algorithm, and not a shared global variable!
+        if (100.0 * total_load_running / total_load_init <=
+            opt::load_ratio_threshold) {
+            VERB0("No more load to balance in system, stopping.");
+            break;
+        }
+

     }
     VERB0("Going to finalize...");
     finalize();
     }
     VERB0("Going to finalize...");
     finalize();


@@ -95,32 +140,46 @@ int process::run()
      */
 
     VERB0("Done.");
      */
 
     VERB0("Done.");

+    INFO3("Final load after %d iteration%s: %g", iter, ESSE(iter), load);

     if (opt::bookkeeping)
     if (opt::bookkeeping)

-        INFO4("Final load after %d iteration%s: %g ; expected: %g",
-              iter, ESSE(iter), load, expected_load);
-    else
-        INFO3("Final load after %d iteration%s: %g", iter, ESSE(iter), load);
+        INFO1("Expected load: %g", expected_load);

     return 0;
 }
 
     return 0;
 }
 

-void process::compute()
+double process::sum_of_to_send() const

 {
 {

-    // fixme: shall we do something special when duration is 0 ?
-    double duration = opt::comp_cost(load);
-    m_task_t task = MSG_task_create("computation", duration, 0.0, NULL);
-    DEBUG2("compute %g flop%s.", duration, ESSE(duration));
-    MSG_task_execute(task);
-    MSG_task_destroy(task);
+    using namespace std::tr1;
+    using namespace std::tr1::placeholders;
+
+    return std::accumulate(neigh.begin(), neigh.end(), 0.0,
+                           bind(std::plus<double>(),
+                                _1, bind(&neighbor::get_to_send, _2)));

 }
 
 double process::load_balance(double /*my_load*/)
 {
 }
 
 double process::load_balance(double /*my_load*/)
 {

+    if (iter == 1)
+        WARN0("process::load_balance is a no-op!");

     return 0.0;
 }
 
     return 0.0;
 }
 

+void process::compute()
+{
+    if (load > 0.0) {
+        double duration = opt::comp_cost(load);
+        m_task_t task = MSG_task_create("computation", duration, 0.0, NULL);
+        DEBUG2("compute %g flop%s.", duration, ESSE(duration));
+        MSG_task_execute(task);
+        MSG_task_destroy(task);
+    } else {
+        DEBUG0("nothing to compute !");
+    }
+}
+

 void process::send1_no_bookkeeping(neighbor& nb)
 {
 void process::send1_no_bookkeeping(neighbor& nb)
 {

-    comm.send(nb.get_ctrl_mbox(), new message(message::INFO, load));
+    if (load != prev_load_broadcast)
+        comm.send(nb.get_ctrl_mbox(), new message(message::INFO, load));

     double load_to_send = nb.get_to_send();
     if (load_to_send > 0.0) {
         comm.send(nb.get_data_mbox(), new message(message::LOAD, load_to_send));
     double load_to_send = nb.get_to_send();
     if (load_to_send > 0.0) {
         comm.send(nb.get_data_mbox(), new message(message::LOAD, load_to_send));


@@ -130,7 +189,9 @@ void process::send1_no_bookkeeping(neighbor& nb)
 
 void process::send1_bookkeeping(neighbor& nb)
 {
 
 void process::send1_bookkeeping(neighbor& nb)
 {

-    comm.send(nb.get_ctrl_mbox(), new message(message::INFO, expected_load));
+    if (expected_load != prev_load_broadcast)
+        comm.send(nb.get_ctrl_mbox(),
+                  new message(message::INFO, expected_load));

     double load_to_send;
     double new_debt;
     double debt_to_send = nb.get_to_send();
     double load_to_send;
     double new_debt;
     double debt_to_send = nb.get_to_send();


@@ -157,28 +218,28 @@ void process::send1_bookkeeping(neighbor& nb)
 
 void process::send()
 {
 
 void process::send()
 {

-    // fixme: shall we send data at all iterations?
+    using namespace std::tr1;
+    using namespace std::tr1::placeholders;
+

     if (opt::bookkeeping) {
         std::for_each(neigh.begin(), neigh.end(),
     if (opt::bookkeeping) {
         std::for_each(neigh.begin(), neigh.end(),

-                      std::tr1::bind(&process::send1_bookkeeping,
-                                     this, std::tr1::placeholders::_1));
+                      bind(&process::send1_bookkeeping, this, _1));
+        prev_load_broadcast = expected_load;

     } else {
         std::for_each(neigh.begin(), neigh.end(),
     } else {
         std::for_each(neigh.begin(), neigh.end(),

-                      std::tr1::bind(&process::send1_no_bookkeeping,
-                                     this, std::tr1::placeholders::_1));
+                      bind(&process::send1_no_bookkeeping, this, _1));
+        prev_load_broadcast = load;

     }
 }
 
     }
 }
 

-// Returns false if a CLOSE message was received. 
-bool process::receive(bool wait_for_close)
+void process::receive(recv_wait_mode wait)

 {
 {

-    bool result = true;
+    // DEBUG1("go for receive(%s)",
+    //        "NO_WAIT\0WAIT\0\0\0\0WAIT_FOR_CLOSE" + 8 * wait);

     message* msg;
     m_host_t from;
     message* msg;
     m_host_t from;

-    while ((ctrl_close_pending ||
-            data_close_pending) && comm.recv(msg, from, wait_for_close)) {
-        DEBUG2("received %s from %s",
-               msg->to_string().c_str(), MSG_host_get_name(from));
+    bool do_wait = (wait != NO_WAIT);
+    while (may_receive && comm.recv(msg, from, do_wait)) {

         switch (msg->get_type()) {
         case message::INFO: {
             neighbor* n = rev_neigh[from];
         switch (msg->get_type()) {
         case message::INFO: {
             neighbor* n = rev_neigh[from];


@@ -188,72 +249,83 @@ bool process::receive(bool wait_for_close)
         case message::CREDIT:
             expected_load += msg->get_amount();
             break;
         case message::CREDIT:
             expected_load += msg->get_amount();
             break;

-        case message::LOAD:
-            load += msg->get_amount();
+        case message::LOAD: {
+            double ld = msg->get_amount();
+            load += ld;
+            if (finalizing)
+                total_load_running -= ld;

             break;
             break;

+        }

         case message::CTRL_CLOSE:
             if (--ctrl_close_pending == 1)
                 comm.next_close_on_ctrl_is_last();
         case message::CTRL_CLOSE:
             if (--ctrl_close_pending == 1)
                 comm.next_close_on_ctrl_is_last();

-            DEBUG1("ctrl_close_pending = %d", ctrl_close_pending);
-            result = false;
+            // DEBUG1("ctrl_close_pending = %d", ctrl_close_pending);
+            close_received = true;

             break;
         case message::DATA_CLOSE:
             if (--data_close_pending == 1)
                 comm.next_close_on_data_is_last();
             break;
         case message::DATA_CLOSE:
             if (--data_close_pending == 1)
                 comm.next_close_on_data_is_last();

-            DEBUG1("data_close_pending = %d", data_close_pending);
-            result = false;
+            // DEBUG1("data_close_pending = %d", data_close_pending);
+            close_received = true;

             break;
         }
         delete msg;
             break;
         }
         delete msg;

+        may_receive = (ctrl_close_pending || data_close_pending);
+        do_wait = (wait == WAIT_FOR_CLOSE);

     }
     }

-    return result;

 }
 
 void process::finalize1(neighbor& nb)
 {
     comm.send(nb.get_ctrl_mbox(), new message(message::CTRL_CLOSE, 0.0));
 }
 
 void process::finalize1(neighbor& nb)
 {
     comm.send(nb.get_ctrl_mbox(), new message(message::CTRL_CLOSE, 0.0));

-    comm.send(nb.get_data_mbox(), new message(message::DATA_CLOSE, 0.0));    
+    comm.send(nb.get_data_mbox(), new message(message::DATA_CLOSE, 0.0));

 }
 
 void process::finalize()
 {
 }
 
 void process::finalize()
 {

+    using namespace std::tr1;
+    using namespace std::tr1::placeholders;
+
+    finalizing = true;
+    total_load_running -= load;
+

     DEBUG2("send CLOSE to %d neighbor%s.",
            (int )neigh.size(), ESSE(neigh.size()));
     std::for_each(neigh.begin(), neigh.end(),
     DEBUG2("send CLOSE to %d neighbor%s.",
            (int )neigh.size(), ESSE(neigh.size()));
     std::for_each(neigh.begin(), neigh.end(),

-                  std::tr1::bind(&process::finalize1,
-                                 this, std::tr1::placeholders::_1));
+                  bind(&process::finalize1, this, _1));

 
     DEBUG2("wait for CLOSE from %d neighbor%s.",
            (int )neigh.size(), ESSE(neigh.size()));
 
     DEBUG2("wait for CLOSE from %d neighbor%s.",
            (int )neigh.size(), ESSE(neigh.size()));

-    receive(true);
+    receive(WAIT_FOR_CLOSE);

 
     comm.flush(true);
 }
 
 
     comm.flush(true);
 }
 

-void process::print_loads(e_xbt_log_priority_t logp)
+#define print_loads_generic(vec, verbose, logp, cat)                    \
+    if (_XBT_LOG_ISENABLEDV((*cat), logp)) {                            \
+        using namespace std::tr1;                                       \
+        using namespace std::tr1::placeholders;                         \
+        XCLOG0(cat, logp, "Neighbor loads:");                           \
+        std::for_each(vec.begin(), vec.end(),                           \
+                      bind(&neighbor::print, _1, verbose, logp, cat));  \
+    } else ((void)0)
+
+void process::print_loads(bool verbose,
+                          e_xbt_log_priority_t logp,
+                          xbt_log_category_t cat) const

 {
 {

-    if (!LOG_ISENABLED(logp))
-        return;
-
-    std::ostringstream oss;
-    if (neigh.empty()) {
-        oss << "no neighbor!";
-    } else {
-        std::transform(neigh.begin(), neigh.end() - 1,
-                       std::ostream_iterator<double>(oss, ", "),
-                       std::tr1::mem_fn(&neighbor::get_load));
-        oss << neigh.back().get_load();
-    }
-    LOG1(logp, "Neighbor loads: %s", oss.str().c_str());
+    print_loads_generic(neigh, verbose, logp, cat);

 }
 
 }
 

-void process::insert_neighbor_in_map(neighbor& nb)
+void process::print_loads_p(bool verbose,
+                            e_xbt_log_priority_t logp,
+                            xbt_log_category_t cat) const

 {
 {

-    neighbor* nbp = &nb;
-    m_host_t host = MSG_get_host_by_name(nb.get_name());
-    rev_neigh.insert(std::make_pair(host, nbp));
+    print_loads_generic(pneigh, verbose, logp, cat);

 }
 
 }
 

+#undef print_loads_generic
+

 // Local variables:
 // mode: c++
 // End:
 // Local variables:
 // mode: c++
 // End: