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"
16 #include <boost/algorithm/string/classification.hpp>
17 #include <boost/algorithm/string/split.hpp>
19 SIMGRID_REGISTER_PLUGIN(host_energy, "Cpu energy consumption.", &sg_host_energy_plugin_init)
21 /** @defgroup plugin_host_energy plugin_host_energy Plugin Host Energy
25 This is the energy plugin, enabling to account not only for computation time, but also for the dissipated energy in the
27 To activate this plugin, first call :cpp:func:`sg_host_energy_plugin_init()` before your loading your platform, and
28 then use :cpp:func:`sg_host_get_consumed_energy()` to retrieve the consumption of a given host.
30 When the host is on, this energy consumption naturally depends on both the current CPU load and the host energy profile.
31 According to our measurements, the consumption is somehow linear in the amount of cores at full speed, with an
32 abnormality when all the cores are idle. The full details are in `our scientific paper
33 <https://hal.inria.fr/hal-01523608>`_ on that topic.
35 As a result, our energy model takes 4 parameters:
37 - ``Idle`` wattage (i.e., instantaneous consumption in Watt) when your host is up and running, but without anything to
39 - ``Epsilon`` wattage when all cores are at 0 or epsilon%, but not in Idle state.
40 - ``AllCores`` wattage when all cores of the host are at 100%.
41 - ``Off`` wattage when the host is turned off.
43 Here is an example of XML declaration:
47 <host id="HostA" speed="100.0Mf" core="4">
48 <prop id="wattage_per_state" value="100.0:120.0:200.0" />
49 <prop id="wattage_off" value="10" />
52 If only two values are given, ``Idle`` is used for the missing ``Epsilon`` value.
54 This example gives the following parameters: ``Off`` is 10 Watts; ``Idle`` is 100 Watts; ``Epsilon`` is 120 Watts and
55 ``AllCores`` is 200 Watts.
56 This is enough to compute the wattage as a function of the amount of loaded cores:
61 <tr><th>#Cores loaded</th><th>Wattage</th><th>Explanation</th></tr>
62 <tr><td>0 (idle)</td><td> 100 Watts </td><td>Idle value</td></tr>
63 <tr><td>1</td><td> 140 Watts</td><td> Linear extrapolation between Epsilon and AllCores</td></tr>
64 <tr><td>2</td><td> 160 Watts</td><td> Linear extrapolation between Epsilon and AllCores</td></tr>
65 <tr><td>3</td><td> 180 Watts</td><td> Linear extrapolation between Epsilon and AllCores</td></tr>
66 <tr><td>4</td><td> 200 Watts</td><td> AllCores value</td></tr>
69 Here is how it looks graphically:
71 .. image:: img/plugin-energy.svg
75 As you can see, the ``Epsilon`` parameter allows to freely specify the slope you want, while using the 2 parameters
76 version of the model (with only ``Idle`` and ``AllCores``) requires that the ``Idle`` value is on the extension of the
77 line crossing the consumption you mesure for each core amount. Please note that specifying the consumption for each core
78 amount separately was not a solution because parallel tasks can use an amount of cores that is not an integer. The good
79 news is that it was not necessary, as our experiments (detailed in the paper) show that the proposed linear model is
80 sufficient to capture reality.
84 <h4>How does DVFS interact with the host energy model?</h4>
86 If your host has several DVFS levels (several pstates), then you should give the energetic profile of each pstate level:
90 <host id="HostC" speed="100.0Mf,50.0Mf,20.0Mf" core="4">
91 <prop id="wattage_per_state"
92 value="95.0:120.0:200.0, 93.0:115.0:170.0, 90.0:110.0:150.0" />
93 <prop id="wattage_off" value="10" />
96 This encodes the following values:
101 <tr><th>pstate</th><th>Performance</th><th>Idle</th><th>Epsilon</th><th>AllCores</th></tr>
102 <tr><td>0</td><td>100 Mflop/s</td><td>95 Watts</td><td>120 Watts</td><td>200 Watts</td></tr>
103 <tr><td>1</td><td>50 Mflop/s</td><td>93 Watts</td><td>115 Watts</td><td>170 Watts</td></tr>
104 <tr><td>2</td><td>20 Mflop/s</td><td>90 Watts</td><td>110 Watts</td><td>150 Watts</td></tr>
107 To change the pstate of a given CPU, use the following functions:
108 :cpp:func:`sg_host_get_nb_pstates()`, :cpp:func:`simgrid::s4u::Host::set_pstate()`,
109 :cpp:func:`sg_host_get_pstate_speed()`.
113 <h4>How accurate are these models?</h4>
115 This model cannot be more accurate than your instantiation: with the default values, your result will not be accurate at
116 all. You can still get accurate energy prediction, provided that you carefully instantiate the model.
117 The first step is to ensure that your timing prediction match perfectly. But this is only the first step of the path,
118 and you really want to read `this paper <https://hal.inria.fr/hal-01523608>`_ to see all what you need to do
119 before you can get accurate energy predictions.
124 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(host_energy, kernel, "Logging specific to the host energy plugin");
126 // Forwards declaration needed to make this function a friend (because friends have external linkage by default)
127 static void on_simulation_end();
129 namespace simgrid::plugin {
138 PowerRange(double idle, double epsilon, double max) : idle_(idle), epsilon_(epsilon), max_(max), slope_(max-epsilon) {}
142 simgrid::s4u::Host* host_ = nullptr;
143 /*< List of (idle_power, epsilon_power, max_power) tuple corresponding to each cpu pstate */
144 std::vector<PowerRange> power_range_watts_list_;
145 bool has_pstate_power_values_ = false; /*< Whether power consumption values were provided for all pstates */
147 /* We need to keep track of what pstate has been used, as we will sometimes be notified only *after* a pstate has been
148 * used (but we need to update the energy consumption with the old pstate!)
151 const int pstate_off_ = -1;
152 double watts_off_ = 0.0; /*< Consumption when the machine is turned off (shutdown) */
153 double total_energy_ = 0.0; /*< Total energy consumed by the host */
154 double last_updated_ = simgrid::s4u::Engine::get_clock(); /*< Timestamp of the last energy update event*/
156 /* Only used to split total energy into unused/used hosts.
157 * If you want to get this info for something else, rather use the host_load plugin
159 bool host_was_used_ = false;
161 void init_watts_range_list();
162 friend void ::on_simulation_end(); // For access to host_was_used_
165 static simgrid::xbt::Extension<simgrid::s4u::Host, HostEnergy> EXTENSION_ID;
167 explicit HostEnergy(simgrid::s4u::Host* ptr);
170 bool has_pstate_power_values() const;
172 double get_current_watts_value();
173 double get_current_watts_value(double cpu_load) const;
174 double get_consumed_energy();
175 double get_watt_idle_at(int pstate) const;
176 double get_watt_min_at(int pstate) const;
177 double get_watt_max_at(int pstate) const;
178 double get_power_range_slope_at(int pstate) const;
179 double get_last_update_time() const { return last_updated_; }
183 simgrid::xbt::Extension<simgrid::s4u::Host, HostEnergy> HostEnergy::EXTENSION_ID;
185 /* Returns whether power consumption values were provided for all pstates. */
186 bool HostEnergy::has_pstate_power_values() const {
187 return has_pstate_power_values_;
190 /* Computes the consumption so far. Called lazily on need. */
191 void HostEnergy::update()
193 double start_time = last_updated_;
194 double finish_time = simgrid::s4u::Engine::get_clock();
196 // We may have start == finish if the past consumption was updated since the simcall was started
197 // for example if 2 actors requested to update the same host's consumption in a given scheduling round.
199 // Even in this case, we need to save the pstate for the next call (after this if),
200 // which may have changed since that recent update.
201 if (start_time < finish_time) {
202 double previous_energy = total_energy_;
204 double instantaneous_power_consumption = this->get_current_watts_value();
206 double energy_this_step = instantaneous_power_consumption * (finish_time - start_time);
208 // TODO Trace: Trace energy_this_step from start_time to finish_time in host->getName()
210 total_energy_ = previous_energy + energy_this_step;
211 last_updated_ = finish_time;
213 XBT_DEBUG("[update_energy of %s] period=[%.8f-%.8f]; current speed=%.2E flop/s (pstate %i); total consumption "
214 "before: %.8f J -> added now: %.8f J",
215 host_->get_cname(), start_time, finish_time, host_->get_pstate_speed(pstate_), pstate_, previous_energy,
219 /* Save data for the upcoming time interval: whether it's on/off and the pstate if it's on */
220 pstate_ = host_->is_on() ? host_->get_pstate() : pstate_off_;
223 HostEnergy::HostEnergy(simgrid::s4u::Host* ptr) : host_(ptr)
225 init_watts_range_list();
227 const char* off_power_str = host_->get_property("wattage_off");
228 if (off_power_str != nullptr) {
230 this->watts_off_ = std::stod(off_power_str);
231 } catch (const std::invalid_argument&) {
232 throw std::invalid_argument("Invalid value for property wattage_off of host " + host_->get_name() + ": " +
236 /* watts_off is 0 by default */
239 HostEnergy::~HostEnergy() = default;
241 double HostEnergy::get_watt_idle_at(int pstate) const
243 if (not has_pstate_power_values_)
245 return power_range_watts_list_[pstate].idle_;
248 double HostEnergy::get_watt_min_at(int pstate) const
250 if (not has_pstate_power_values_)
252 return power_range_watts_list_[pstate].epsilon_;
255 double HostEnergy::get_watt_max_at(int pstate) const
257 if (not has_pstate_power_values_)
259 return power_range_watts_list_[pstate].max_;
262 double HostEnergy::get_power_range_slope_at(int pstate) const
264 if (not has_pstate_power_values_)
266 return power_range_watts_list_[pstate].slope_;
269 /** @brief Computes the power consumed by the host according to the current situation
271 * - If the host is off, that's the watts_off value
272 * - if it's on, take the current pstate and the current processor load into account */
273 double HostEnergy::get_current_watts_value()
275 if (this->pstate_ == pstate_off_) // The host is off (or was off at the beginning of this time interval)
276 return this->watts_off_;
278 double current_speed = host_->get_pstate_speed(this->pstate_);
282 if (current_speed <= 0)
283 // Some users declare a pstate of speed 0 flops (e.g., to model boot time).
284 // We consider that the machine is then fully loaded. That's arbitrary but it avoids a NaN
287 cpu_load = host_->get_load() / current_speed;
289 /* Divide by the number of cores here to have a value between 0 and 1 */
290 cpu_load /= host_->get_core_count();
292 if (cpu_load > 1) // This condition is true for energy_ptask on 32 bits, even if cpu_load is displayed as 1.000000
293 cpu_load = 1; // That may be an harmless rounding error?
295 host_was_used_ = true;
298 return get_current_watts_value(cpu_load);
301 /** @brief Computes the power that the host would consume at the provided processor load
303 * Whether the host is ON or OFF is not taken into account.
305 double HostEnergy::get_current_watts_value(double cpu_load) const
307 if (not has_pstate_power_values_)
310 /* Return watts_off if pstate == pstate_off (ie, if the host is off) */
311 if (this->pstate_ == pstate_off_) {
315 PowerRange power_range = power_range_watts_list_.at(this->pstate_);
316 double current_power;
321 * Something is going on, the host is not idle.
323 * The power consumption follows the regular model:
324 * P(cpu_load) = Pstatic + Pdynamic * cpu_load
325 * where Pstatic = power_range.epsilon_ and Pdynamic = power_range.slope_
326 * and the cpu_load is a value between 0 and 1.
328 current_power = power_range.epsilon_ + cpu_load * power_range.slope_;
332 /* The host is idle, take the dedicated value! */
333 current_power = power_range.idle_;
336 XBT_DEBUG("[get_current_watts] pstate=%i, epsilon_power=%f, max_power=%f, slope=%f", this->pstate_, power_range.epsilon_,
337 power_range.max_, power_range.slope_);
338 XBT_DEBUG("[get_current_watts] Current power (watts) = %f, load = %f", current_power, cpu_load);
340 return current_power;
343 double HostEnergy::get_consumed_energy()
345 if (last_updated_ < simgrid::s4u::Engine::get_clock()) // We need to simcall this as it modifies the environment
346 simgrid::kernel::actor::simcall_answered(std::bind(&HostEnergy::update, this));
348 return total_energy_;
351 void HostEnergy::init_watts_range_list()
353 const char* all_power_values_str = host_->get_property("wattage_per_state");
354 if (all_power_values_str == nullptr) {
355 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());
359 std::vector<std::string> all_power_values;
360 boost::split(all_power_values, all_power_values_str, boost::is_any_of(","));
361 XBT_DEBUG("%s: power properties: %s", host_->get_cname(), all_power_values_str);
363 xbt_assert(all_power_values.size() == host_->get_pstate_count(),
364 "Invalid XML file. Found %zu energetic profiles for %lu pstates", all_power_values.size(),
365 host_->get_pstate_count());
368 for (auto const& current_power_values_str : all_power_values) {
369 /* retrieve the power values associated with the pstate i */
370 std::vector<std::string> current_power_values;
371 boost::split(current_power_values, current_power_values_str, boost::is_any_of(":"));
373 xbt_assert(current_power_values.size() == 2 || current_power_values.size() == 3,
374 "Power properties incorrectly defined for host %s."
375 "It should be 'Idle:AllCores' (or 'Idle:Epsilon:AllCores') power values.",
379 double epsilon_power;
382 auto msg_idle = xbt::string_printf("Invalid Idle value for pstate %d on host %s", i, host_->get_cname());
383 auto msg_epsilon = xbt::string_printf("Invalid Epsilon value for pstate %d on host %s", i, host_->get_cname());
384 auto msg_max = xbt::string_printf("Invalid AllCores value for pstate %d on host %s", i, host_->get_cname());
386 idle_power = xbt_str_parse_double((current_power_values.at(0)).c_str(), msg_idle.c_str());
387 if (current_power_values.size() == 2) { // Case: Idle:AllCores
388 epsilon_power = xbt_str_parse_double((current_power_values.at(0)).c_str(), msg_idle.c_str());
389 max_power = xbt_str_parse_double((current_power_values.at(1)).c_str(), msg_max.c_str());
390 } else { // Case: Idle:Epsilon:AllCores
391 epsilon_power = xbt_str_parse_double((current_power_values.at(1)).c_str(), msg_epsilon.c_str());
392 max_power = xbt_str_parse_double((current_power_values.at(2)).c_str(), msg_max.c_str());
395 XBT_DEBUG("Creating PowerRange for host %s. Idle:%f, Epsilon:%f, AllCores:%f.", host_->get_cname(), idle_power, epsilon_power, max_power);
397 PowerRange range(idle_power, epsilon_power, max_power);
398 power_range_watts_list_.push_back(range);
402 has_pstate_power_values_ = true;
404 } // namespace simgrid::plugin
406 using simgrid::plugin::HostEnergy;
408 /* **************************** events callback *************************** */
409 static void on_creation(simgrid::s4u::Host& host)
411 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
414 // TODO Trace: set to zero the energy variable associated to host->get_name()
416 host.extension_set(new HostEnergy(&host));
419 static void on_action_state_change(simgrid::kernel::resource::CpuAction const& action,
420 simgrid::kernel::resource::Action::State /*previous*/)
422 for (simgrid::kernel::resource::CpuImpl* const& cpu : action.cpus()) {
423 simgrid::s4u::Host* host = cpu->get_iface();
424 if (host != nullptr) {
425 // If it's a VM, take the corresponding PM
426 if (const auto* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host))
429 // Get the host_energy extension for the relevant host
430 auto* host_energy = host->extension<HostEnergy>();
432 if (host_energy->get_last_update_time() < simgrid::s4u::Engine::get_clock())
433 host_energy->update();
438 /* This callback is fired either when the host changes its state (on/off) ("onStateChange") or its speed
439 * (because the user changed the pstate, or because of external trace events) ("onSpeedChange") */
440 static void on_host_change(simgrid::s4u::Host const& host)
442 if (dynamic_cast<simgrid::s4u::VirtualMachine const*>(&host)) // Ignore virtual machines
445 auto* host_energy = host.extension<HostEnergy>();
447 host_energy->update();
450 static void on_host_destruction(simgrid::s4u::Host const& host)
452 if (dynamic_cast<simgrid::s4u::VirtualMachine const*>(&host)) // Ignore virtual machines
455 XBT_INFO("Energy consumption of host %s: %f Joules", host.get_cname(),
456 host.extension<HostEnergy>()->get_consumed_energy());
459 static void on_simulation_end()
461 double total_energy = 0.0; // Total energy consumption (whole platform)
462 double used_hosts_energy = 0.0; // Energy consumed by hosts that computed something
463 for (simgrid::s4u::Host const* host : simgrid::s4u::Engine::get_instance()->get_all_hosts()) {
464 if (host && dynamic_cast<const simgrid::s4u::VirtualMachine*>(host) == nullptr) { // Ignore virtual machines
465 double energy = host->extension<HostEnergy>()->get_consumed_energy();
466 total_energy += energy;
467 if (host->extension<HostEnergy>()->host_was_used_)
468 used_hosts_energy += energy;
471 XBT_INFO("Total energy consumption: %f Joules (used hosts: %f Joules; unused/idle hosts: %f)", total_energy,
472 used_hosts_energy, total_energy - used_hosts_energy);
475 /* **************************** Public interface *************************** */
477 /** @ingroup plugin_host_energy
478 * @brief Enable host energy plugin
479 * @details Enable energy plugin to get joules consumption of each cpu. Call this function before loading your platform.
481 void sg_host_energy_plugin_init()
483 if (HostEnergy::EXTENSION_ID.valid())
486 HostEnergy::EXTENSION_ID = simgrid::s4u::Host::extension_create<HostEnergy>();
488 simgrid::s4u::Host::on_creation_cb(&on_creation);
489 simgrid::s4u::Host::on_state_change_cb(&on_host_change);
490 simgrid::s4u::Host::on_speed_change_cb(&on_host_change);
491 simgrid::s4u::Host::on_destruction_cb(&on_host_destruction);
492 simgrid::s4u::Engine::on_simulation_end_cb(&on_simulation_end);
493 simgrid::kernel::resource::CpuAction::on_state_change.connect(&on_action_state_change);
494 // We may only have one actor on a node. If that actor executes something like
495 // compute -> recv -> compute
496 // the recv operation will not trigger a "CpuAction::on_state_change". This means
497 // that the next trigger would be the 2nd compute, hence ignoring the idle time
498 // during the recv call. By updating at the beginning of a compute, we can
499 // fix that. (If the cpu is not idle, this is not required.)
500 simgrid::s4u::Exec::on_start_cb([](simgrid::s4u::Exec const& activity) {
501 if (activity.get_host_number() == 1) { // We only run on one host
502 simgrid::s4u::Host* host = activity.get_host();
503 if (const auto* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host))
505 xbt_assert(host != nullptr);
506 host->extension<HostEnergy>()->update();
511 /** @ingroup plugin_host_energy
512 * @brief updates the consumption of all hosts
514 * After this call, sg_host_get_consumed_energy() will not interrupt your process
515 * (until after the next clock update).
517 void sg_host_energy_update_all()
519 simgrid::kernel::actor::simcall_answered([]() {
520 std::vector<simgrid::s4u::Host*> list = simgrid::s4u::Engine::get_instance()->get_all_hosts();
521 for (auto const& host : list)
522 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(host) == nullptr) { // Ignore virtual machines
523 xbt_assert(host != nullptr);
524 host->extension<HostEnergy>()->update();
529 static void ensure_plugin_inited()
531 if (not HostEnergy::EXTENSION_ID.valid())
532 throw simgrid::xbt::InitializationError("The Energy plugin is not active. Please call sg_host_energy_plugin_init() "
533 "before calling any function related to that plugin.");
536 /** @ingroup plugin_host_energy
537 * @brief Returns the total energy consumed by the host so far (in Joules)
539 * Please note that since the consumption is lazily updated, it may require a simcall to update it.
540 * The result is that the actor requesting this value will be interrupted,
541 * the value will be updated in kernel mode before returning the control to the requesting actor.
543 double sg_host_get_consumed_energy(const_sg_host_t host)
545 ensure_plugin_inited();
546 auto host_energy = host->extension<HostEnergy>();
547 xbt_assert(host_energy->has_pstate_power_values(), "No power range properties specified for host %s",
549 return host_energy->get_consumed_energy();
552 /** @ingroup plugin_host_energy
553 * @brief Get the amount of watt dissipated when the host is idling
555 double sg_host_get_idle_consumption(const_sg_host_t host)
557 ensure_plugin_inited();
558 return host->extension<HostEnergy>()->get_watt_idle_at(0);
561 /** @ingroup plugin_host_energy
562 * @brief Get the amount of watt dissipated at the given pstate when the host is idling
564 double sg_host_get_idle_consumption_at(const_sg_host_t host, int pstate)
566 ensure_plugin_inited();
567 return host->extension<HostEnergy>()->get_watt_idle_at(pstate);
570 /** @ingroup plugin_host_energy
571 * @brief Get the amount of watt dissipated at the given pstate when the host is at 0 or epsilon% CPU usage.
573 double sg_host_get_wattmin_at(const_sg_host_t host, int pstate)
575 ensure_plugin_inited();
576 return host->extension<HostEnergy>()->get_watt_min_at(pstate);
578 /** @ingroup plugin_host_energy
579 * @brief Returns the amount of watt dissipated at the given pstate when the host burns CPU at 100%
581 double sg_host_get_wattmax_at(const_sg_host_t host, int pstate)
583 ensure_plugin_inited();
584 return host->extension<HostEnergy>()->get_watt_max_at(pstate);
586 /** @ingroup plugin_host_energy
587 * @brief Returns the power slope at the given pstate
589 double sg_host_get_power_range_slope_at(const_sg_host_t host, int pstate)
591 ensure_plugin_inited();
592 return host->extension<HostEnergy>()->get_power_range_slope_at(pstate);
594 /** @ingroup plugin_host_energy
595 * @brief Returns the current consumption of the host
597 double sg_host_get_current_consumption(const_sg_host_t host)
599 ensure_plugin_inited();
600 auto host_energy = host->extension<HostEnergy>();
601 xbt_assert(host_energy->has_pstate_power_values(), "No power range properties specified for host %s",
603 return host_energy->get_current_watts_value();