Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
Apply the default settings of 'smpi/buffering' too
[simgrid.git] / src / plugins / host_dvfs.cpp
index c4d4d70..0290509 100644 (file)
@@ -1,4 +1,4 @@
-/* Copyright (c) 2010-2018. The SimGrid Team. All rights reserved.          */
+/* Copyright (c) 2010-2019. The SimGrid Team. All rights reserved.          */
 
 /* This program is free software; you can redistribute it and/or modify it
  * under the terms of the license (GNU LGPL) which comes with this package. */
@@ -6,12 +6,18 @@
 #include "simgrid/plugins/dvfs.h"
 #include "simgrid/plugins/load.h"
 #include "simgrid/s4u/Engine.hpp"
+#include "src/internal_config.h" // HAVE_SMPI
 #include "src/kernel/activity/ExecImpl.hpp"
 #include "src/plugins/vm/VirtualMachineImpl.hpp"
+#if HAVE_SMPI
 #include "src/smpi/plugins/ampi/ampi.hpp"
+#endif
 #include <xbt/config.hpp>
 
 #include <boost/algorithm/string.hpp>
+#if HAVE_SMPI
+#include "src/smpi/include/smpi_request.hpp"
+#endif
 
 SIMGRID_REGISTER_PLUGIN(host_dvfs, "Dvfs support", &sg_host_dvfs_plugin_init)
 
@@ -20,26 +26,30 @@ static simgrid::config::Flag<double> cfg_sampling_rate("plugin/dvfs/sampling-rat
     [](double val){if (val != 0.1) sg_host_dvfs_plugin_init();});
 
 static simgrid::config::Flag<std::string> cfg_governor("plugin/dvfs/governor",
-    "Which Governor should be used that adapts the CPU frequency?", "performance",
-
-    std::map<std::string, std::string>({
-        {"adagio", "TODO: Doc"},
-        {"conservative", "TODO: Doc"},
-        {"ondemand", "TODO: Doc"},
-        {"performance", "TODO: Doc"},
-        {"powersave", "TODO: Doc"},
-    }),
-
-    [](std::string val) { if (val != "performance") sg_host_dvfs_plugin_init(); });
-
-static simgrid::config::Flag<int> cfg_min_pstate("plugin/dvfs/min-pstate", {"plugin/dvfs/min_pstate"},
-    "Which pstate is the minimum (and hence fastest) pstate for this governor?", 0,
-    [](int index) {});
+                                                       "Which Governor should be used that adapts the CPU frequency?",
+                                                       "performance",
+
+                                                       std::map<std::string, std::string>({
+#if HAVE_SMPI
+                                                         {"adagio", "TODO: Doc"},
+#endif
+                                                             {"conservative", "TODO: Doc"}, {"ondemand", "TODO: Doc"},
+                                                             {"performance", "TODO: Doc"}, {"powersave", "TODO: Doc"},
+                                                       }),
+
+                                                       [](const std::string& val) {
+                                                         if (val != "performance")
+                                                           sg_host_dvfs_plugin_init();
+                                                       });
+
+static simgrid::config::Flag<int>
+    cfg_min_pstate("plugin/dvfs/min-pstate", {"plugin/dvfs/min_pstate"},
+                   "Which pstate is the minimum (and hence fastest) pstate for this governor?", 0);
 
 static const int max_pstate_not_limited = -1;
-static simgrid::config::Flag<int> cfg_max_pstate("plugin/dvfs/max-pstate", {"plugin/dvfs/max_pstate"},
-    "Which pstate is the maximum (and hence slowest) pstate for this governor?", max_pstate_not_limited,
-    [](int index) {});
+static simgrid::config::Flag<int>
+    cfg_max_pstate("plugin/dvfs/max-pstate", {"plugin/dvfs/max_pstate"},
+                   "Which pstate is the maximum (and hence slowest) pstate for this governor?", max_pstate_not_limited);
 
 /** @addtogroup SURF_plugin_load
 
@@ -77,8 +87,8 @@ class Governor {
 private:
   simgrid::s4u::Host* const host_;
   double sampling_rate_;
-  int min_pstate;
-  int max_pstate;
+  int min_pstate; //< Never use a pstate less than this one
+  int max_pstate; //< Never use a pstate larger than this one
 
 public:
   explicit Governor(simgrid::s4u::Host* ptr)
@@ -91,6 +101,8 @@ public:
   virtual ~Governor() = default;
   virtual std::string get_name() const = 0;
   simgrid::s4u::Host* get_host() const { return host_; }
+  int get_min_pstate() const { return min_pstate; }
+  int get_max_pstate() const { return max_pstate; }
 
   void init()
   {
@@ -133,7 +145,7 @@ public:
   explicit Performance(simgrid::s4u::Host* ptr) : Governor(ptr) {}
   std::string get_name() const override { return "Performance"; }
 
-  void update() override { get_host()->set_pstate(min_pstate); }
+  void update() override { get_host()->set_pstate(get_min_pstate()); }
 };
 
 /**
@@ -151,7 +163,7 @@ public:
   explicit Powersave(simgrid::s4u::Host* ptr) : Governor(ptr) {}
   std::string get_name() const override { return "Powersave"; }
 
-  void update() override { get_host()->set_pstate(max_pstate); }
+  void update() override { get_host()->set_pstate(get_max_pstate()); }
 };
 
 /**
@@ -180,8 +192,8 @@ public:
     sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
 
     if (load > freq_up_threshold_) {
-      get_host()->set_pstate(min_pstate); /* Run at max. performance! */
-      XBT_INFO("Load: %f > threshold: %f --> changed to pstate %i", load, freq_up_threshold_, min_pstate);
+      get_host()->set_pstate(get_min_pstate()); /* Run at max. performance! */
+      XBT_INFO("Load: %f > threshold: %f --> changed to pstate %i", load, freq_up_threshold_, get_min_pstate());
     } else {
       /* The actual implementation uses a formula here: (See Kernel file cpufreq_ondemand.c:158)
        *
@@ -192,13 +204,14 @@ public:
        */
       // Load is now < freq_up_threshold; exclude pstate 0 (the fastest)
       // because pstate 0 can only be selected if load > freq_up_threshold_
-      int new_pstate = max_pstate - load * (max_pstate + 1);
+      int new_pstate = get_max_pstate() - load * (get_max_pstate() + 1);
+      if (new_pstate < get_min_pstate())
+        new_pstate = get_min_pstate();
       get_host()->set_pstate(new_pstate);
 
       XBT_DEBUG("Load: %f < threshold: %f --> changed to pstate %i", load, freq_up_threshold_, new_pstate);
     }
   }
-
 };
 
 /**
@@ -228,7 +241,7 @@ public:
     sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
 
     if (load > freq_up_threshold_) {
-      if (pstate != min_pstate) {
+      if (pstate != get_min_pstate()) {
         get_host()->set_pstate(pstate - 1);
         XBT_INFO("Load: %f > threshold: %f -> increasing performance to pstate %d", load, freq_up_threshold_,
                  pstate - 1);
@@ -237,7 +250,7 @@ public:
                   freq_up_threshold_, pstate);
       }
     } else if (load < freq_down_threshold_) {
-      if (pstate != max_pstate) { // Are we in the slowest pstate already?
+      if (pstate != get_max_pstate()) { // Are we in the slowest pstate already?
         get_host()->set_pstate(pstate + 1);
         XBT_INFO("Load: %f < threshold: %f -> slowing down to pstate %d", load, freq_down_threshold_, pstate + 1);
       } else {
@@ -248,6 +261,7 @@ public:
   }
 };
 
+#if HAVE_SMPI
 class Adagio : public Governor {
 private:
   int best_pstate     = 0;
@@ -255,43 +269,47 @@ private:
   double comp_counter = 0;
   double comp_timer   = 0;
 
-  std::vector<std::vector<double>> rates;
+  std::vector<std::vector<double>> rates; // Each host + all frequencies of that host
 
   unsigned int task_id   = 0;
   bool iteration_running = false; /*< Are we currently between iteration_in and iteration_out calls? */
 
 public:
   explicit Adagio(simgrid::s4u::Host* ptr)
-      : Governor(ptr), rates(100, std::vector<double>(host_->get_pstate_count(), 0.0))
+      : Governor(ptr), rates(100, std::vector<double>(ptr->get_pstate_count(), 0.0))
   {
-    simgrid::smpi::plugin::ampi::on_iteration_in.connect([this](simgrid::s4u::ActorPtr actor) {
+    simgrid::smpi::plugin::ampi::on_iteration_in.connect([this](simgrid::s4u::Actor const& actor) {
       // Every instance of this class subscribes to this event, so one per host
       // This means that for any actor, all 'hosts' are normally notified of these
       // changes, even those who don't currently run the actor 'proc_id'.
       // -> Let's check if this signal call is for us!
-      if (get_host() == actor->get_host()) {
+      if (get_host() == actor.get_host()) {
         iteration_running = true;
       }
     });
-    simgrid::smpi::plugin::ampi::on_iteration_out.connect([this](simgrid::s4u::ActorPtr actor) {
-      if (get_host() == actor->get_host()) {
+    simgrid::smpi::plugin::ampi::on_iteration_out.connect([this](simgrid::s4u::Actor const& actor) {
+      if (get_host() == actor.get_host()) {
         iteration_running = false;
         task_id           = 0;
       }
     });
-    simgrid::kernel::activity::ExecImpl::on_creation.connect([this](simgrid::kernel::activity::ExecImplPtr activity) {
-      if (activity->host_ == get_host())
-        pre_task();
-    });
-    simgrid::kernel::activity::ExecImpl::on_completion.connect([this](simgrid::kernel::activity::ExecImplPtr activity) {
-      // For more than one host (not yet supported), we can access the host via
-      // simcalls_.front()->issuer->iface()->get_host()
-      if (activity->host_ == get_host() && iteration_running) {
-        comp_timer += activity->surf_action_->get_finish_time() - activity->surf_action_->get_start_time();
-      }
-    });
+    simgrid::kernel::activity::ExecImpl::on_creation.connect(
+        [this](simgrid::kernel::activity::ExecImpl const& activity) {
+          if (activity.get_host() == get_host())
+            pre_task();
+        });
+    simgrid::kernel::activity::ExecImpl::on_completion.connect(
+        [this](simgrid::kernel::activity::ExecImpl const& activity) {
+          // For more than one host (not yet supported), we can access the host via
+          // simcalls_.front()->issuer->iface()->get_host()
+          if (activity.get_host() == get_host() && iteration_running) {
+            comp_timer += activity.surf_action_->get_finish_time() - activity.surf_action_->get_start_time();
+          }
+        });
+    // FIXME I think that this fires at the same time for all hosts, so when the src sends something,
+    // the dst will be notified even though it didn't even arrive at the recv yet
     simgrid::s4u::Link::on_communicate.connect(
-        [this](kernel::resource::NetworkAction* action, s4u::Host* src, s4u::Host* dst) {
+        [this](kernel::resource::NetworkAction const&, s4u::Host* src, s4u::Host* dst) {
           if ((get_host() == src || get_host() == dst) && iteration_running) {
             post_task();
           }
@@ -302,36 +320,34 @@ public:
 
   void pre_task()
   {
-    sg_host_load_reset(host_);
-    comp_counter = sg_host_get_computed_flops(host_); // Should be 0 because of the reset
+    sg_host_load_reset(get_host());
+    comp_counter = sg_host_get_computed_flops(get_host()); // Should be 0 because of the reset
     comp_timer   = 0;
     start_time   = simgrid::s4u::Engine::get_clock();
     if (rates.size() <= task_id)
-      rates.resize(task_id + 5, std::vector<double>(host_->get_pstate_count(), 0.0));
+      rates.resize(task_id + 5, std::vector<double>(get_host()->get_pstate_count(), 0.0));
     if (rates[task_id][best_pstate] == 0)
       best_pstate = 0;
-    host_->set_pstate(best_pstate); // Load our schedule
+    get_host()->set_pstate(best_pstate); // Load our schedule
     XBT_DEBUG("Set pstate to %i", best_pstate);
   }
 
   void post_task()
   {
-    double computed_flops = sg_host_get_computed_flops(host_) - comp_counter;
+    double computed_flops = sg_host_get_computed_flops(get_host()) - comp_counter;
     double target_time    = (simgrid::s4u::Engine::get_clock() - start_time);
-    target_time =
-        target_time *
-        static_cast<double>(99.0 / 100.0); // FIXME We account for t_copy arbitrarily with 1% -- this needs to be fixed
+    target_time           = target_time * 99.0 / 100.0; // FIXME We account for t_copy arbitrarily with 1%
+                                                        // -- this needs to be fixed
 
     bool is_initialized         = rates[task_id][best_pstate] != 0;
     rates[task_id][best_pstate] = computed_flops / comp_timer;
     if (not is_initialized) {
-      for (int i = 1; i < host_->get_pstate_count(); i++) {
-        rates[task_id][i] = rates[task_id][0] * (host_->get_pstate_speed(i) / host_->get_speed());
+      for (int i = 1; i < get_host()->get_pstate_count(); i++) {
+        rates[task_id][i] = rates[task_id][0] * (get_host()->get_pstate_speed(i) / get_host()->get_speed());
       }
-      is_initialized = true;
     }
 
-    for (int pstate = host_->get_pstate_count() - 1; pstate >= 0; pstate--) {
+    for (int pstate = get_host()->get_pstate_count() - 1; pstate >= 0; pstate--) {
       if (computed_flops / rates[task_id][pstate] <= target_time) {
         // We just found the pstate we want to use!
         best_pstate = pstate;
@@ -343,6 +359,7 @@ public:
 
   virtual void update() override {}
 };
+#endif
 } // namespace dvfs
 } // namespace plugin
 } // namespace simgrid
@@ -381,10 +398,14 @@ static void on_host_added(simgrid::s4u::Host& host)
       } else if (dvfs_governor == "ondemand") {
         return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
             new simgrid::plugin::dvfs::OnDemand(daemon_proc->get_host()));
-      } else if (dvfs_governor == "adagio") {
+      }
+#if HAVE_SMPI
+      else if (dvfs_governor == "adagio") {
         return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
             new simgrid::plugin::dvfs::Adagio(daemon_proc->get_host()));
-      } else if (dvfs_governor == "performance") {
+      }
+#endif
+      else if (dvfs_governor == "performance") {
         return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
             new simgrid::plugin::dvfs::Performance(daemon_proc->get_host()));
       } else if (dvfs_governor == "powersave") {