1 /* Copyright (c) 2010-2023. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
6 #include <simgrid/Exception.hpp>
7 #include <simgrid/plugins/energy.h>
8 #include <simgrid/s4u/Engine.hpp>
9 #include <simgrid/s4u/Exec.hpp>
10 #include <simgrid/s4u/Host.hpp>
11 #include <simgrid/s4u/VirtualMachine.hpp>
12 #include <simgrid/simix.hpp>
14 #include "src/kernel/resource/CpuImpl.hpp"
15 #include "src/simgrid/module.hpp"
17 #include <boost/algorithm/string/classification.hpp>
18 #include <boost/algorithm/string/split.hpp>
20 SIMGRID_REGISTER_PLUGIN(host_energy, "Cpu energy consumption.", &sg_host_energy_plugin_init)
22 /** @defgroup plugin_host_energy plugin_host_energy Plugin Host Energy
26 This is the energy plugin, enabling to account not only for computation time, but also for the dissipated energy in the
28 To activate this plugin, first call :cpp:func:`sg_host_energy_plugin_init()` before your loading your platform, and
29 then use :cpp:func:`sg_host_get_consumed_energy()` to retrieve the consumption of a given host.
31 When the host is on, this energy consumption naturally depends on both the current CPU load and the host energy profile.
32 According to our measurements, the consumption is somehow linear in the amount of cores at full speed, with an
33 abnormality when all the cores are idle. The full details are in `our scientific paper
34 <https://hal.inria.fr/hal-01523608>`_ on that topic.
36 As a result, our energy model takes 4 parameters:
38 - ``Idle`` wattage (i.e., instantaneous consumption in Watt) when your host is up and running, but without anything to
40 - ``Epsilon`` wattage when all cores are at 0 or epsilon%, but not in Idle state.
41 - ``AllCores`` wattage when all cores of the host are at 100%.
42 - ``Off`` wattage when the host is turned off.
44 Here is an example of XML declaration:
48 <host id="HostA" speed="100.0Mf" core="4">
49 <prop id="wattage_per_state" value="100.0:120.0:200.0" />
50 <prop id="wattage_off" value="10" />
53 If only two values are given, ``Idle`` is used for the missing ``Epsilon`` value.
55 This example gives the following parameters: ``Off`` is 10 Watts; ``Idle`` is 100 Watts; ``Epsilon`` is 120 Watts and
56 ``AllCores`` is 200 Watts.
57 This is enough to compute the wattage as a function of the amount of loaded cores:
62 <tr><th>#Cores loaded</th><th>Wattage</th><th>Explanation</th></tr>
63 <tr><td>0 (idle)</td><td> 100 Watts </td><td>Idle value</td></tr>
64 <tr><td>1</td><td> 140 Watts</td><td> Linear extrapolation between Epsilon and AllCores</td></tr>
65 <tr><td>2</td><td> 160 Watts</td><td> Linear extrapolation between Epsilon and AllCores</td></tr>
66 <tr><td>3</td><td> 180 Watts</td><td> Linear extrapolation between Epsilon and AllCores</td></tr>
67 <tr><td>4</td><td> 200 Watts</td><td> AllCores value</td></tr>
70 Here is how it looks graphically:
72 .. image:: img/plugin-energy.svg
76 As you can see, the ``Epsilon`` parameter allows to freely specify the slope you want, while using the 2 parameters
77 version of the model (with only ``Idle`` and ``AllCores``) requires that the ``Idle`` value is on the extension of the
78 line crossing the consumption you mesure for each core amount. Please note that specifying the consumption for each core
79 amount separately was not a solution because parallel tasks can use an amount of cores that is not an integer. The good
80 news is that it was not necessary, as our experiments (detailed in the paper) show that the proposed linear model is
81 sufficient to capture reality.
85 <h4>How does DVFS interact with the host energy model?</h4>
87 If your host has several DVFS levels (several pstates), then you should give the energetic profile of each pstate level:
91 <host id="HostC" speed="100.0Mf,50.0Mf,20.0Mf" core="4">
92 <prop id="wattage_per_state"
93 value="95.0:120.0:200.0, 93.0:115.0:170.0, 90.0:110.0:150.0" />
94 <prop id="wattage_off" value="10" />
97 This encodes the following values:
102 <tr><th>pstate</th><th>Performance</th><th>Idle</th><th>Epsilon</th><th>AllCores</th></tr>
103 <tr><td>0</td><td>100 Mflop/s</td><td>95 Watts</td><td>120 Watts</td><td>200 Watts</td></tr>
104 <tr><td>1</td><td>50 Mflop/s</td><td>93 Watts</td><td>115 Watts</td><td>170 Watts</td></tr>
105 <tr><td>2</td><td>20 Mflop/s</td><td>90 Watts</td><td>110 Watts</td><td>150 Watts</td></tr>
108 To change the pstate of a given CPU, use the following functions:
109 :cpp:func:`sg_host_get_nb_pstates()`, :cpp:func:`simgrid::s4u::Host::set_pstate()`,
110 :cpp:func:`sg_host_get_pstate_speed()`.
114 <h4>How accurate are these models?</h4>
116 This model cannot be more accurate than your instantiation: with the default values, your result will not be accurate at
117 all. You can still get accurate energy prediction, provided that you carefully instantiate the model.
118 The first step is to ensure that your timing prediction match perfectly. But this is only the first step of the path,
119 and you really want to read `this paper <https://hal.inria.fr/hal-01523608>`_ to see all what you need to do
120 before you can get accurate energy predictions.
125 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(host_energy, kernel, "Logging specific to the host energy plugin");
127 // Forwards declaration needed to make this function a friend (because friends have external linkage by default)
128 static void on_simulation_end();
130 namespace simgrid::plugin {
139 PowerRange(double idle, double epsilon, double max) : idle_(idle), epsilon_(epsilon), max_(max), slope_(max-epsilon) {}
143 simgrid::s4u::Host* host_ = nullptr;
144 /*< List of (idle_power, epsilon_power, max_power) tuple corresponding to each cpu pstate */
145 std::vector<PowerRange> power_range_watts_list_;
146 bool has_pstate_power_values_ = false; /*< Whether power consumption values were provided for all pstates */
148 /* We need to keep track of what pstate has been used, as we will sometimes be notified only *after* a pstate has been
149 * used (but we need to update the energy consumption with the old pstate!)
152 const int pstate_off_ = -1;
153 double watts_off_ = 0.0; /*< Consumption when the machine is turned off (shutdown) */
154 double total_energy_ = 0.0; /*< Total energy consumed by the host */
155 double last_updated_ = simgrid::s4u::Engine::get_clock(); /*< Timestamp of the last energy update event*/
157 /* Only used to split total energy into unused/used hosts.
158 * If you want to get this info for something else, rather use the host_load plugin
160 bool host_was_used_ = false;
162 void init_watts_range_list();
163 friend void ::on_simulation_end(); // For access to host_was_used_
166 static simgrid::xbt::Extension<simgrid::s4u::Host, HostEnergy> EXTENSION_ID;
168 explicit HostEnergy(simgrid::s4u::Host* ptr);
171 bool has_pstate_power_values() const;
173 double get_current_watts_value();
174 double get_current_watts_value(double cpu_load) const;
175 double get_consumed_energy();
176 double get_watt_idle_at(int pstate) const;
177 double get_watt_min_at(int pstate) const;
178 double get_watt_max_at(int pstate) const;
179 double get_power_range_slope_at(int pstate) const;
180 double get_last_update_time() const { return last_updated_; }
184 simgrid::xbt::Extension<simgrid::s4u::Host, HostEnergy> HostEnergy::EXTENSION_ID;
186 /* Returns whether power consumption values were provided for all pstates. */
187 bool HostEnergy::has_pstate_power_values() const {
188 return has_pstate_power_values_;
191 /* Computes the consumption so far. Called lazily on need. */
192 void HostEnergy::update()
194 double start_time = last_updated_;
195 double finish_time = simgrid::s4u::Engine::get_clock();
197 // We may have start == finish if the past consumption was updated since the simcall was started
198 // for example if 2 actors requested to update the same host's consumption in a given scheduling round.
200 // Even in this case, we need to save the pstate for the next call (after this if),
201 // which may have changed since that recent update.
202 if (start_time < finish_time) {
203 double previous_energy = total_energy_;
205 double instantaneous_power_consumption = this->get_current_watts_value();
207 double energy_this_step = instantaneous_power_consumption * (finish_time - start_time);
209 // TODO Trace: Trace energy_this_step from start_time to finish_time in host->getName()
211 total_energy_ = previous_energy + energy_this_step;
212 last_updated_ = finish_time;
214 XBT_DEBUG("[update_energy of %s] period=[%.8f-%.8f]; current speed=%.2E flop/s (pstate %i); total consumption "
215 "before: %.8f J -> added now: %.8f J",
216 host_->get_cname(), start_time, finish_time, host_->get_pstate_speed(pstate_), pstate_, previous_energy,
220 /* Save data for the upcoming time interval: whether it's on/off and the pstate if it's on */
221 pstate_ = host_->is_on() ? host_->get_pstate() : pstate_off_;
224 HostEnergy::HostEnergy(simgrid::s4u::Host* ptr) : host_(ptr)
226 init_watts_range_list();
228 const char* off_power_str = host_->get_property("wattage_off");
229 if (off_power_str != nullptr) {
231 this->watts_off_ = std::stod(off_power_str);
232 } catch (const std::invalid_argument&) {
233 throw std::invalid_argument("Invalid value for property wattage_off of host " + host_->get_name() + ": " +
237 /* watts_off is 0 by default */
240 HostEnergy::~HostEnergy() = default;
242 double HostEnergy::get_watt_idle_at(int pstate) const
244 if (not has_pstate_power_values_)
246 return power_range_watts_list_[pstate].idle_;
249 double HostEnergy::get_watt_min_at(int pstate) const
251 if (not has_pstate_power_values_)
253 return power_range_watts_list_[pstate].epsilon_;
256 double HostEnergy::get_watt_max_at(int pstate) const
258 if (not has_pstate_power_values_)
260 return power_range_watts_list_[pstate].max_;
263 double HostEnergy::get_power_range_slope_at(int pstate) const
265 if (not has_pstate_power_values_)
267 return power_range_watts_list_[pstate].slope_;
270 /** @brief Computes the power consumed by the host according to the current situation
272 * - If the host is off, that's the watts_off value
273 * - if it's on, take the current pstate and the current processor load into account */
274 double HostEnergy::get_current_watts_value()
276 if (this->pstate_ == pstate_off_) // The host is off (or was off at the beginning of this time interval)
277 return this->watts_off_;
279 double current_speed = host_->get_pstate_speed(this->pstate_);
283 if (current_speed <= 0)
284 // Some users declare a pstate of speed 0 flops (e.g., to model boot time).
285 // We consider that the machine is then fully loaded. That's arbitrary but it avoids a NaN
288 cpu_load = host_->get_load() / current_speed;
290 /* Divide by the number of cores here to have a value between 0 and 1 */
291 cpu_load /= host_->get_core_count();
293 if (cpu_load > 1) // This condition is true for energy_ptask on 32 bits, even if cpu_load is displayed as 1.000000
294 cpu_load = 1; // That may be an harmless rounding error?
296 host_was_used_ = true;
299 return get_current_watts_value(cpu_load);
302 /** @brief Computes the power that the host would consume at the provided processor load
304 * Whether the host is ON or OFF is not taken into account.
306 double HostEnergy::get_current_watts_value(double cpu_load) const
308 if (not has_pstate_power_values_)
311 /* Return watts_off if pstate == pstate_off (ie, if the host is off) */
312 if (this->pstate_ == pstate_off_) {
316 PowerRange power_range = power_range_watts_list_.at(this->pstate_);
317 double current_power;
322 * Something is going on, the host is not idle.
324 * The power consumption follows the regular model:
325 * P(cpu_load) = Pstatic + Pdynamic * cpu_load
326 * where Pstatic = power_range.epsilon_ and Pdynamic = power_range.slope_
327 * and the cpu_load is a value between 0 and 1.
329 current_power = power_range.epsilon_ + cpu_load * power_range.slope_;
333 /* The host is idle, take the dedicated value! */
334 current_power = power_range.idle_;
337 XBT_DEBUG("[get_current_watts] pstate=%i, epsilon_power=%f, max_power=%f, slope=%f", this->pstate_, power_range.epsilon_,
338 power_range.max_, power_range.slope_);
339 XBT_DEBUG("[get_current_watts] Current power (watts) = %f, load = %f", current_power, cpu_load);
341 return current_power;
344 double HostEnergy::get_consumed_energy()
346 if (last_updated_ < simgrid::s4u::Engine::get_clock()) // We need to simcall this as it modifies the environment
347 simgrid::kernel::actor::simcall_answered(std::bind(&HostEnergy::update, this));
349 return total_energy_;
352 void HostEnergy::init_watts_range_list()
354 const char* all_power_values_str = host_->get_property("wattage_per_state");
355 if (all_power_values_str == nullptr) {
356 XBT_WARN("No energetic profiles (wattage_per_state) given for host %s, using 0 W by default. Direct request of power/energy consumption of this host will fail.", host_->get_cname());
360 std::vector<std::string> all_power_values;
361 boost::split(all_power_values, all_power_values_str, boost::is_any_of(","));
362 XBT_DEBUG("%s: power properties: %s", host_->get_cname(), all_power_values_str);
364 xbt_assert(all_power_values.size() == host_->get_pstate_count(),
365 "Invalid XML file. Found %zu energetic profiles for %lu pstates", all_power_values.size(),
366 host_->get_pstate_count());
369 for (auto const& current_power_values_str : all_power_values) {
370 /* retrieve the power values associated with the pstate i */
371 std::vector<std::string> current_power_values;
372 boost::split(current_power_values, current_power_values_str, boost::is_any_of(":"));
374 xbt_assert(current_power_values.size() == 2 || current_power_values.size() == 3,
375 "Power properties incorrectly defined for host %s."
376 "It should be 'Idle:AllCores' (or 'Idle:Epsilon:AllCores') power values.",
380 double epsilon_power;
383 auto msg_idle = xbt::string_printf("Invalid Idle value for pstate %d on host %s", i, host_->get_cname());
384 auto msg_epsilon = xbt::string_printf("Invalid Epsilon value for pstate %d on host %s", i, host_->get_cname());
385 auto msg_max = xbt::string_printf("Invalid AllCores value for pstate %d on host %s", i, host_->get_cname());
387 idle_power = xbt_str_parse_double((current_power_values.at(0)).c_str(), msg_idle.c_str());
388 if (current_power_values.size() == 2) { // Case: Idle:AllCores
389 epsilon_power = xbt_str_parse_double((current_power_values.at(0)).c_str(), msg_idle.c_str());
390 max_power = xbt_str_parse_double((current_power_values.at(1)).c_str(), msg_max.c_str());
391 } else { // Case: Idle:Epsilon:AllCores
392 epsilon_power = xbt_str_parse_double((current_power_values.at(1)).c_str(), msg_epsilon.c_str());
393 max_power = xbt_str_parse_double((current_power_values.at(2)).c_str(), msg_max.c_str());
396 XBT_DEBUG("Creating PowerRange for host %s. Idle:%f, Epsilon:%f, AllCores:%f.", host_->get_cname(), idle_power, epsilon_power, max_power);
398 PowerRange range(idle_power, epsilon_power, max_power);
399 power_range_watts_list_.push_back(range);
403 has_pstate_power_values_ = true;
405 } // namespace simgrid::plugin
407 using simgrid::plugin::HostEnergy;
409 /* **************************** events callback *************************** */
410 static void on_creation(simgrid::s4u::Host& host)
412 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
415 // TODO Trace: set to zero the energy variable associated to host->get_name()
417 host.extension_set(new HostEnergy(&host));
420 static void on_action_state_change(simgrid::kernel::resource::CpuAction const& action,
421 simgrid::kernel::resource::Action::State /*previous*/)
423 for (simgrid::kernel::resource::CpuImpl* const& cpu : action.cpus()) {
424 simgrid::s4u::Host* host = cpu->get_iface();
425 if (host != nullptr) {
426 // If it's a VM, take the corresponding PM
427 if (const auto* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host))
430 // Get the host_energy extension for the relevant host
431 auto* host_energy = host->extension<HostEnergy>();
433 if (host_energy->get_last_update_time() < simgrid::s4u::Engine::get_clock())
434 host_energy->update();
439 /* This callback is fired either when the host changes its state (on/off) ("onStateChange") or its speed
440 * (because the user changed the pstate, or because of external trace events) ("onSpeedChange") */
441 static void on_host_change(simgrid::s4u::Host const& host)
443 if (dynamic_cast<simgrid::s4u::VirtualMachine const*>(&host)) // Ignore virtual machines
446 auto* host_energy = host.extension<HostEnergy>();
448 host_energy->update();
451 static void on_host_destruction(simgrid::s4u::Host const& host)
453 if (dynamic_cast<simgrid::s4u::VirtualMachine const*>(&host)) // Ignore virtual machines
456 XBT_INFO("Energy consumption of host %s: %f Joules", host.get_cname(),
457 host.extension<HostEnergy>()->get_consumed_energy());
460 static void on_simulation_end()
462 double total_energy = 0.0; // Total energy consumption (whole platform)
463 double used_hosts_energy = 0.0; // Energy consumed by hosts that computed something
464 for (simgrid::s4u::Host const* host : simgrid::s4u::Engine::get_instance()->get_all_hosts()) {
465 if (host && dynamic_cast<const simgrid::s4u::VirtualMachine*>(host) == nullptr) { // Ignore virtual machines
466 double energy = host->extension<HostEnergy>()->get_consumed_energy();
467 total_energy += energy;
468 if (host->extension<HostEnergy>()->host_was_used_)
469 used_hosts_energy += energy;
472 XBT_INFO("Total energy consumption: %f Joules (used hosts: %f Joules; unused/idle hosts: %f)", total_energy,
473 used_hosts_energy, total_energy - used_hosts_energy);
476 /* **************************** Public interface *************************** */
478 /** @ingroup plugin_host_energy
479 * @brief Enable host energy plugin
480 * @details Enable energy plugin to get joules consumption of each cpu. Call this function before loading your platform.
482 void sg_host_energy_plugin_init()
484 if (HostEnergy::EXTENSION_ID.valid())
487 HostEnergy::EXTENSION_ID = simgrid::s4u::Host::extension_create<HostEnergy>();
489 simgrid::s4u::Host::on_creation_cb(&on_creation);
490 simgrid::s4u::Host::on_state_change_cb(&on_host_change);
491 simgrid::s4u::Host::on_speed_change_cb(&on_host_change);
492 simgrid::s4u::Host::on_destruction_cb(&on_host_destruction);
493 simgrid::s4u::Engine::on_simulation_end_cb(&on_simulation_end);
494 simgrid::kernel::resource::CpuAction::on_state_change.connect(&on_action_state_change);
495 // We may only have one actor on a node. If that actor executes something like
496 // compute -> recv -> compute
497 // the recv operation will not trigger a "CpuAction::on_state_change". This means
498 // that the next trigger would be the 2nd compute, hence ignoring the idle time
499 // during the recv call. By updating at the beginning of a compute, we can
500 // fix that. (If the cpu is not idle, this is not required.)
501 simgrid::s4u::Exec::on_start_cb([](simgrid::s4u::Exec const& activity) {
502 if (activity.get_host_number() == 1) { // We only run on one host
503 simgrid::s4u::Host* host = activity.get_host();
504 if (const auto* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host))
506 xbt_assert(host != nullptr);
507 host->extension<HostEnergy>()->update();
512 /** @ingroup plugin_host_energy
513 * @brief updates the consumption of all hosts
515 * After this call, sg_host_get_consumed_energy() will not interrupt your process
516 * (until after the next clock update).
518 void sg_host_energy_update_all()
520 simgrid::kernel::actor::simcall_answered([]() {
521 std::vector<simgrid::s4u::Host*> list = simgrid::s4u::Engine::get_instance()->get_all_hosts();
522 for (auto const& host : list)
523 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(host) == nullptr) { // Ignore virtual machines
524 xbt_assert(host != nullptr);
525 host->extension<HostEnergy>()->update();
530 static void ensure_plugin_inited()
532 if (not HostEnergy::EXTENSION_ID.valid())
533 throw simgrid::xbt::InitializationError("The Energy plugin is not active. Please call sg_host_energy_plugin_init() "
534 "before calling any function related to that plugin.");
537 /** @ingroup plugin_host_energy
538 * @brief Returns the total energy consumed by the host so far (in Joules)
540 * Please note that since the consumption is lazily updated, it may require a simcall to update it.
541 * The result is that the actor requesting this value will be interrupted,
542 * the value will be updated in kernel mode before returning the control to the requesting actor.
544 double sg_host_get_consumed_energy(const_sg_host_t host)
546 ensure_plugin_inited();
547 auto host_energy = host->extension<HostEnergy>();
548 xbt_assert(host_energy->has_pstate_power_values(), "No power range properties specified for host %s",
550 return host_energy->get_consumed_energy();
553 /** @ingroup plugin_host_energy
554 * @brief Get the amount of watt dissipated when the host is idling
556 double sg_host_get_idle_consumption(const_sg_host_t host)
558 ensure_plugin_inited();
559 return host->extension<HostEnergy>()->get_watt_idle_at(0);
562 /** @ingroup plugin_host_energy
563 * @brief Get the amount of watt dissipated at the given pstate when the host is idling
565 double sg_host_get_idle_consumption_at(const_sg_host_t host, int pstate)
567 ensure_plugin_inited();
568 return host->extension<HostEnergy>()->get_watt_idle_at(pstate);
571 /** @ingroup plugin_host_energy
572 * @brief Get the amount of watt dissipated at the given pstate when the host is at 0 or epsilon% CPU usage.
574 double sg_host_get_wattmin_at(const_sg_host_t host, int pstate)
576 ensure_plugin_inited();
577 return host->extension<HostEnergy>()->get_watt_min_at(pstate);
579 /** @ingroup plugin_host_energy
580 * @brief Returns the amount of watt dissipated at the given pstate when the host burns CPU at 100%
582 double sg_host_get_wattmax_at(const_sg_host_t host, int pstate)
584 ensure_plugin_inited();
585 return host->extension<HostEnergy>()->get_watt_max_at(pstate);
587 /** @ingroup plugin_host_energy
588 * @brief Returns the power slope at the given pstate
590 double sg_host_get_power_range_slope_at(const_sg_host_t host, int pstate)
592 ensure_plugin_inited();
593 return host->extension<HostEnergy>()->get_power_range_slope_at(pstate);
595 /** @ingroup plugin_host_energy
596 * @brief Returns the current consumption of the host
598 double sg_host_get_current_consumption(const_sg_host_t host)
600 ensure_plugin_inited();
601 auto host_energy = host->extension<HostEnergy>();
602 xbt_assert(host_energy->has_pstate_power_values(), "No power range properties specified for host %s",
604 return host_energy->get_current_watts_value();