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/activity/ActivityImpl.hpp"
15 #include "src/kernel/resource/CpuImpl.hpp"
16 #include "src/simgrid/module.hpp"
18 #include <boost/algorithm/string/classification.hpp>
19 #include <boost/algorithm/string/split.hpp>
21 SIMGRID_REGISTER_PLUGIN(host_energy, "Cpu energy consumption.", &sg_host_energy_plugin_init)
23 /** @defgroup plugin_host_energy plugin_host_energy Plugin Host Energy
27 This is the energy plugin, enabling to account not only for computation time, but also for the dissipated energy in the
29 To activate this plugin, first call :cpp:func:`sg_host_energy_plugin_init()` before your loading your platform, and
30 then use :cpp:func:`sg_host_get_consumed_energy()` to retrieve the consumption of a given host.
32 When the host is on, this energy consumption naturally depends on both the current CPU load and the host energy profile.
33 According to our measurements, the consumption is somehow linear in the amount of cores at full speed, with an
34 abnormality when all the cores are idle. The full details are in `our scientific paper
35 <https://hal.inria.fr/hal-01523608>`_ on that topic.
37 As a result, our energy model takes 4 parameters:
39 - ``Idle`` wattage (i.e., instantaneous consumption in Watt) when your host is up and running, but without anything to
41 - ``Epsilon`` wattage when all cores are at 0 or epsilon%, but not in Idle state.
42 - ``AllCores`` wattage when all cores of the host are at 100%.
43 - ``Off`` wattage when the host is turned off.
45 Here is an example of XML declaration:
49 <host id="HostA" speed="100.0Mf" core="4">
50 <prop id="wattage_per_state" value="100.0:120.0:200.0" />
51 <prop id="wattage_off" value="10" />
54 If only two values are given, ``Idle`` is used for the missing ``Epsilon`` value.
56 This example gives the following parameters: ``Off`` is 10 Watts; ``Idle`` is 100 Watts; ``Epsilon`` is 120 Watts and
57 ``AllCores`` is 200 Watts.
58 This is enough to compute the wattage as a function of the amount of loaded cores:
63 <tr><th>#Cores loaded</th><th>Wattage</th><th>Explanation</th></tr>
64 <tr><td>0 (idle)</td><td> 100 Watts </td><td>Idle value</td></tr>
65 <tr><td>1</td><td> 140 Watts</td><td> Linear extrapolation between Epsilon and AllCores</td></tr>
66 <tr><td>2</td><td> 160 Watts</td><td> Linear extrapolation between Epsilon and AllCores</td></tr>
67 <tr><td>3</td><td> 180 Watts</td><td> Linear extrapolation between Epsilon and AllCores</td></tr>
68 <tr><td>4</td><td> 200 Watts</td><td> AllCores value</td></tr>
71 Here is how it looks graphically:
73 .. image:: img/plugin-energy.svg
77 As you can see, the ``Epsilon`` parameter allows to freely specify the slope you want, while using the 2 parameters
78 version of the model (with only ``Idle`` and ``AllCores``) requires that the ``Idle`` value is on the extension of the
79 line crossing the consumption you mesure for each core amount. Please note that specifying the consumption for each core
80 amount separately was not a solution because parallel tasks can use an amount of cores that is not an integer. The good
81 news is that it was not necessary, as our experiments (detailed in the paper) show that the proposed linear model is
82 sufficient to capture reality.
86 <h4>How does DVFS interact with the host energy model?</h4>
88 If your host has several DVFS levels (several pstates), then you should give the energetic profile of each pstate level:
92 <host id="HostC" speed="100.0Mf,50.0Mf,20.0Mf" core="4">
93 <prop id="wattage_per_state"
94 value="95.0:120.0:200.0, 93.0:115.0:170.0, 90.0:110.0:150.0" />
95 <prop id="wattage_off" value="10" />
98 This encodes the following values:
103 <tr><th>pstate</th><th>Performance</th><th>Idle</th><th>Epsilon</th><th>AllCores</th></tr>
104 <tr><td>0</td><td>100 Mflop/s</td><td>95 Watts</td><td>120 Watts</td><td>200 Watts</td></tr>
105 <tr><td>1</td><td>50 Mflop/s</td><td>93 Watts</td><td>115 Watts</td><td>170 Watts</td></tr>
106 <tr><td>2</td><td>20 Mflop/s</td><td>90 Watts</td><td>110 Watts</td><td>150 Watts</td></tr>
109 To change the pstate of a given CPU, use the following functions:
110 :cpp:func:`sg_host_get_nb_pstates()`, :cpp:func:`simgrid::s4u::Host::set_pstate()`,
111 :cpp:func:`sg_host_get_pstate_speed()`.
115 <h4>How accurate are these models?</h4>
117 This model cannot be more accurate than your instantiation: with the default values, your result will not be accurate at
118 all. You can still get accurate energy prediction, provided that you carefully instantiate the model.
119 The first step is to ensure that your timing prediction match perfectly. But this is only the first step of the path,
120 and you really want to read `this paper <https://hal.inria.fr/hal-01523608>`_ to see all what you need to do
121 before you can get accurate energy predictions.
126 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(host_energy, kernel, "Logging specific to the host energy plugin");
128 // Forwards declaration needed to make this function a friend (because friends have external linkage by default)
129 static void on_simulation_end();
131 namespace simgrid::plugin {
140 PowerRange(double idle, double epsilon, double max) : idle_(idle), epsilon_(epsilon), max_(max), slope_(max-epsilon) {}
144 simgrid::s4u::Host* host_ = nullptr;
145 /*< List of (idle_power, epsilon_power, max_power) tuple corresponding to each cpu pstate */
146 std::vector<PowerRange> power_range_watts_list_;
147 bool has_pstate_power_values_ = false; /*< Whether power consumption values were provided for all pstates */
149 /* We need to keep track of what pstate has been used, as we will sometimes be notified only *after* a pstate has been
150 * used (but we need to update the energy consumption with the old pstate!)
153 const int pstate_off_ = -1;
154 double watts_off_ = 0.0; /*< Consumption when the machine is turned off (shutdown) */
155 double total_energy_ = 0.0; /*< Total energy consumed by the host */
156 double last_updated_ = simgrid::s4u::Engine::get_clock(); /*< Timestamp of the last energy update event*/
158 /* Only used to split total energy into unused/used hosts.
159 * If you want to get this info for something else, rather use the host_load plugin
161 bool host_was_used_ = false;
163 void init_watts_range_list();
164 friend void ::on_simulation_end(); // For access to host_was_used_
167 static simgrid::xbt::Extension<simgrid::s4u::Host, HostEnergy> EXTENSION_ID;
169 explicit HostEnergy(simgrid::s4u::Host* ptr);
172 bool has_pstate_power_values() const;
174 double get_current_watts_value();
175 double get_current_watts_value(double cpu_load) const;
176 double get_consumed_energy();
177 double get_watt_idle_at(int pstate) const;
178 double get_watt_min_at(int pstate) const;
179 double get_watt_max_at(int pstate) const;
180 double get_power_range_slope_at(int pstate) const;
181 double get_last_update_time() const { return last_updated_; }
185 simgrid::xbt::Extension<simgrid::s4u::Host, HostEnergy> HostEnergy::EXTENSION_ID;
187 /* Returns whether power consumption values were provided for all pstates. */
188 bool HostEnergy::has_pstate_power_values() const {
189 return has_pstate_power_values_;
192 /* Computes the consumption so far. Called lazily on need. */
193 void HostEnergy::update()
195 double start_time = last_updated_;
196 double finish_time = simgrid::s4u::Engine::get_clock();
198 // We may have start == finish if the past consumption was updated since the simcall was started
199 // for example if 2 actors requested to update the same host's consumption in a given scheduling round.
201 // Even in this case, we need to save the pstate for the next call (after this if),
202 // which may have changed since that recent update.
203 if (start_time < finish_time) {
204 double previous_energy = total_energy_;
206 double instantaneous_power_consumption = this->get_current_watts_value();
208 double energy_this_step = instantaneous_power_consumption * (finish_time - start_time);
210 // TODO Trace: Trace energy_this_step from start_time to finish_time in host->getName()
212 total_energy_ = previous_energy + energy_this_step;
213 last_updated_ = finish_time;
215 XBT_DEBUG("[update_energy of %s] period=[%.8f-%.8f]; current speed=%.2E flop/s (pstate %i); total consumption "
216 "before: %.8f J -> added now: %.8f J",
217 host_->get_cname(), start_time, finish_time, host_->get_pstate_speed(pstate_), pstate_, previous_energy,
221 /* Save data for the upcoming time interval: whether it's on/off and the pstate if it's on */
222 pstate_ = host_->is_on() ? host_->get_pstate() : pstate_off_;
225 HostEnergy::HostEnergy(simgrid::s4u::Host* ptr) : host_(ptr)
227 init_watts_range_list();
229 const char* off_power_str = host_->get_property("wattage_off");
230 if (off_power_str != nullptr) {
232 this->watts_off_ = std::stod(off_power_str);
233 } catch (const std::invalid_argument&) {
234 throw std::invalid_argument("Invalid value for property wattage_off of host " + host_->get_name() + ": " +
238 /* watts_off is 0 by default */
241 HostEnergy::~HostEnergy() = default;
243 double HostEnergy::get_watt_idle_at(int pstate) const
245 if (not has_pstate_power_values_)
247 return power_range_watts_list_[pstate].idle_;
250 double HostEnergy::get_watt_min_at(int pstate) const
252 if (not has_pstate_power_values_)
254 return power_range_watts_list_[pstate].epsilon_;
257 double HostEnergy::get_watt_max_at(int pstate) const
259 if (not has_pstate_power_values_)
261 return power_range_watts_list_[pstate].max_;
264 double HostEnergy::get_power_range_slope_at(int pstate) const
266 if (not has_pstate_power_values_)
268 return power_range_watts_list_[pstate].slope_;
271 /** @brief Computes the power consumed by the host according to the current situation
273 * - If the host is off, that's the watts_off value
274 * - if it's on, take the current pstate and the current processor load into account */
275 double HostEnergy::get_current_watts_value()
277 if (this->pstate_ == pstate_off_) // The host is off (or was off at the beginning of this time interval)
278 return this->watts_off_;
280 double current_speed = host_->get_pstate_speed(this->pstate_);
284 if (current_speed <= 0)
285 // Some users declare a pstate of speed 0 flops (e.g., to model boot time).
286 // We consider that the machine is then fully loaded. That's arbitrary but it avoids a NaN
289 cpu_load = host_->get_load() / current_speed;
291 /* Divide by the number of cores here to have a value between 0 and 1 */
292 cpu_load /= host_->get_core_count();
294 if (cpu_load > 1) // This condition is true for energy_ptask on 32 bits, even if cpu_load is displayed as 1.000000
295 cpu_load = 1; // That may be an harmless rounding error?
297 host_was_used_ = true;
300 return get_current_watts_value(cpu_load);
303 /** @brief Computes the power that the host would consume at the provided processor load
305 * Whether the host is ON or OFF is not taken into account.
307 double HostEnergy::get_current_watts_value(double cpu_load) const
309 if (not has_pstate_power_values_)
312 /* Return watts_off if pstate == pstate_off (ie, if the host is off) */
313 if (this->pstate_ == pstate_off_) {
317 PowerRange power_range = power_range_watts_list_.at(this->pstate_);
318 double current_power;
323 * Something is going on, the host is not idle.
325 * The power consumption follows the regular model:
326 * P(cpu_load) = Pstatic + Pdynamic * cpu_load
327 * where Pstatic = power_range.epsilon_ and Pdynamic = power_range.slope_
328 * and the cpu_load is a value between 0 and 1.
330 current_power = power_range.epsilon_ + cpu_load * power_range.slope_;
334 /* The host is idle, take the dedicated value! */
335 current_power = power_range.idle_;
338 XBT_DEBUG("[get_current_watts] pstate=%i, epsilon_power=%f, max_power=%f, slope=%f", this->pstate_, power_range.epsilon_,
339 power_range.max_, power_range.slope_);
340 XBT_DEBUG("[get_current_watts] Current power (watts) = %f, load = %f", current_power, cpu_load);
342 return current_power;
345 double HostEnergy::get_consumed_energy()
347 if (last_updated_ < simgrid::s4u::Engine::get_clock()) // We need to simcall this as it modifies the environment
348 simgrid::kernel::actor::simcall_answered(std::bind(&HostEnergy::update, this));
350 return total_energy_;
353 void HostEnergy::init_watts_range_list()
355 const char* all_power_values_str = host_->get_property("wattage_per_state");
356 if (all_power_values_str == nullptr) {
357 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());
361 std::vector<std::string> all_power_values;
362 boost::split(all_power_values, all_power_values_str, boost::is_any_of(","));
363 XBT_DEBUG("%s: power properties: %s", host_->get_cname(), all_power_values_str);
365 xbt_assert(all_power_values.size() == host_->get_pstate_count(),
366 "Invalid XML file. Found %zu energetic profiles for %lu pstates", all_power_values.size(),
367 host_->get_pstate_count());
370 for (auto const& current_power_values_str : all_power_values) {
371 /* retrieve the power values associated with the pstate i */
372 std::vector<std::string> current_power_values;
373 boost::split(current_power_values, current_power_values_str, boost::is_any_of(":"));
375 xbt_assert(current_power_values.size() == 2 || current_power_values.size() == 3,
376 "Power properties incorrectly defined for host %s."
377 "It should be 'Idle:AllCores' (or 'Idle:Epsilon:AllCores') power values.",
381 double epsilon_power;
384 auto msg_idle = xbt::string_printf("Invalid Idle value for pstate %d on host %s", i, host_->get_cname());
385 auto msg_epsilon = xbt::string_printf("Invalid Epsilon value for pstate %d on host %s", i, host_->get_cname());
386 auto msg_max = xbt::string_printf("Invalid AllCores value for pstate %d on host %s", i, host_->get_cname());
388 idle_power = xbt_str_parse_double((current_power_values.at(0)).c_str(), msg_idle.c_str());
389 if (current_power_values.size() == 2) { // Case: Idle:AllCores
390 epsilon_power = xbt_str_parse_double((current_power_values.at(0)).c_str(), msg_idle.c_str());
391 max_power = xbt_str_parse_double((current_power_values.at(1)).c_str(), msg_max.c_str());
392 } else { // Case: Idle:Epsilon:AllCores
393 epsilon_power = xbt_str_parse_double((current_power_values.at(1)).c_str(), msg_epsilon.c_str());
394 max_power = xbt_str_parse_double((current_power_values.at(2)).c_str(), msg_max.c_str());
397 XBT_DEBUG("Creating PowerRange for host %s. Idle:%f, Epsilon:%f, AllCores:%f.", host_->get_cname(), idle_power, epsilon_power, max_power);
399 PowerRange range(idle_power, epsilon_power, max_power);
400 power_range_watts_list_.push_back(range);
404 has_pstate_power_values_ = true;
406 } // namespace simgrid::plugin
408 using simgrid::plugin::HostEnergy;
410 /* **************************** events callback *************************** */
411 static void on_creation(simgrid::s4u::Host& host)
413 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
416 // TODO Trace: set to zero the energy variable associated to host->get_name()
418 host.extension_set(new HostEnergy(&host));
421 static void on_action_state_change(simgrid::kernel::resource::CpuAction const& action,
422 simgrid::kernel::resource::Action::State /*previous*/)
424 for (simgrid::kernel::resource::CpuImpl const* cpu : action.cpus()) {
425 simgrid::s4u::Host* host = cpu->get_iface();
426 if (host != nullptr) {
427 // If it's a VM, take the corresponding PM
428 if (const auto* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host))
431 // Get the host_energy extension for the relevant host
432 auto* host_energy = host->extension<HostEnergy>();
434 if (host_energy->get_last_update_time() < simgrid::s4u::Engine::get_clock())
435 host_energy->update();
440 /* This callback is fired either when the host changes its state (on/off) ("onStateChange") or its speed
441 * (because the user changed the pstate, or because of external trace events) ("onSpeedChange") */
442 static void on_host_change(simgrid::s4u::Host const& h)
444 const auto* host = &h;
445 if (const auto* vm = dynamic_cast<simgrid::s4u::VirtualMachine const*>(host)) // Take the PM of virtual machines
448 host->extension<HostEnergy>()->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 static void on_activity_suspend_resume(simgrid::s4u::Activity const& activity)
478 if (const auto* action = dynamic_cast<simgrid::kernel::resource::CpuAction*>(activity.get_impl()->model_action_))
479 on_action_state_change(*action, /*ignored*/ action->get_state());
482 /* **************************** Public interface *************************** */
484 /** @ingroup plugin_host_energy
485 * @brief Enable host energy plugin
486 * @details Enable energy plugin to get joules consumption of each cpu. Call this function before loading your platform.
488 void sg_host_energy_plugin_init()
490 if (HostEnergy::EXTENSION_ID.valid())
493 HostEnergy::EXTENSION_ID = simgrid::s4u::Host::extension_create<HostEnergy>();
495 simgrid::s4u::Host::on_creation_cb(&on_creation);
496 simgrid::s4u::Host::on_onoff_cb(&on_host_change);
497 simgrid::s4u::Host::on_speed_change_cb(&on_host_change);
498 simgrid::s4u::Host::on_destruction_cb(&on_host_destruction);
499 simgrid::s4u::Host::on_exec_state_change_cb(&on_action_state_change);
500 simgrid::s4u::VirtualMachine::on_suspend_cb(&on_host_change);
501 simgrid::s4u::VirtualMachine::on_resume_cb(&on_host_change);
502 simgrid::s4u::Exec::on_suspend_cb(on_activity_suspend_resume);
503 simgrid::s4u::Exec::on_resume_cb(on_activity_suspend_resume);
504 simgrid::s4u::Engine::on_simulation_end_cb(&on_simulation_end);
505 // We may only have one actor on a node. If that actor executes something like
506 // compute -> recv -> compute
507 // the recv operation will not trigger a "Host::on_exec_state_change_cb". This means
508 // that the next trigger would be the 2nd compute, hence ignoring the idle time
509 // during the recv call. By updating at the beginning of a compute, we can
510 // fix that. (If the cpu is not idle, this is not required.)
511 simgrid::s4u::Exec::on_start_cb([](simgrid::s4u::Exec const& activity) {
512 if (activity.get_host_number() == 1) { // We only run on one host
513 simgrid::s4u::Host* host = activity.get_host();
514 if (const auto* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host))
516 xbt_assert(host != nullptr);
517 host->extension<HostEnergy>()->update();
522 /** @ingroup plugin_host_energy
523 * @brief updates the consumption of all hosts
525 * After this call, sg_host_get_consumed_energy() will not interrupt your process
526 * (until after the next clock update).
528 void sg_host_energy_update_all()
530 simgrid::kernel::actor::simcall_answered([]() {
531 std::vector<simgrid::s4u::Host*> list = simgrid::s4u::Engine::get_instance()->get_all_hosts();
532 for (auto const& host : list)
533 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(host) == nullptr) { // Ignore virtual machines
534 xbt_assert(host != nullptr);
535 host->extension<HostEnergy>()->update();
540 static void ensure_plugin_inited()
542 if (not HostEnergy::EXTENSION_ID.valid())
543 throw simgrid::xbt::InitializationError("The Energy plugin is not active. Please call sg_host_energy_plugin_init() "
544 "before calling any function related to that plugin.");
547 /** @ingroup plugin_host_energy
548 * @brief Returns the total energy consumed by the host so far (in Joules)
550 * Please note that since the consumption is lazily updated, it may require a simcall to update it.
551 * The result is that the actor requesting this value will be interrupted,
552 * the value will be updated in kernel mode before returning the control to the requesting actor.
554 double sg_host_get_consumed_energy(const_sg_host_t host)
556 ensure_plugin_inited();
557 auto* host_energy = host->extension<HostEnergy>();
558 xbt_assert(host_energy->has_pstate_power_values(), "No power range properties specified for host %s",
560 return host_energy->get_consumed_energy();
563 /** @ingroup plugin_host_energy
564 * @brief Get the amount of watt dissipated when the host is idling
566 double sg_host_get_idle_consumption(const_sg_host_t host)
568 ensure_plugin_inited();
569 return host->extension<HostEnergy>()->get_watt_idle_at(0);
572 /** @ingroup plugin_host_energy
573 * @brief Get the amount of watt dissipated at the given pstate when the host is idling
575 double sg_host_get_idle_consumption_at(const_sg_host_t host, int pstate)
577 ensure_plugin_inited();
578 return host->extension<HostEnergy>()->get_watt_idle_at(pstate);
581 /** @ingroup plugin_host_energy
582 * @brief Get the amount of watt dissipated at the given pstate when the host is at 0 or epsilon% CPU usage.
584 double sg_host_get_wattmin_at(const_sg_host_t host, int pstate)
586 ensure_plugin_inited();
587 return host->extension<HostEnergy>()->get_watt_min_at(pstate);
589 /** @ingroup plugin_host_energy
590 * @brief Returns the amount of watt dissipated at the given pstate when the host burns CPU at 100%
592 double sg_host_get_wattmax_at(const_sg_host_t host, int pstate)
594 ensure_plugin_inited();
595 return host->extension<HostEnergy>()->get_watt_max_at(pstate);
597 /** @ingroup plugin_host_energy
598 * @brief Returns the power slope at the given pstate
600 double sg_host_get_power_range_slope_at(const_sg_host_t host, int pstate)
602 ensure_plugin_inited();
603 return host->extension<HostEnergy>()->get_power_range_slope_at(pstate);
605 /** @ingroup plugin_host_energy
606 * @brief Returns the current consumption of the host
608 double sg_host_get_current_consumption(const_sg_host_t host)
610 ensure_plugin_inited();
611 auto* host_energy = host->extension<HostEnergy>();
612 xbt_assert(host_energy->has_pstate_power_values(), "No power range properties specified for host %s",
614 return host_energy->get_current_watts_value();