1 /* Copyright (c) 2010-2021. 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 "src/include/surf/surf.hpp"
11 #include "src/kernel/activity/ExecImpl.hpp"
12 #include "src/plugins/vm/VirtualMachineImpl.hpp"
13 #include "src/surf/cpu_interface.hpp"
15 #include <boost/algorithm/string/classification.hpp>
16 #include <boost/algorithm/string/split.hpp>
18 SIMGRID_REGISTER_PLUGIN(host_energy, "Cpu energy consumption.", &sg_host_energy_plugin_init)
20 /** @defgroup plugin_host_energy plugin_host_energy Plugin Host Energy
24 This is the energy plugin, enabling to account not only for computation time, but also for the dissipated energy in the
26 To activate this plugin, first call :cpp:func:`sg_host_energy_plugin_init()` before your loading your platform, and
27 then use :cpp:func:`sg_host_get_consumed_energy()` to retrieve the consumption of a given host.
29 When the host is on, this energy consumption naturally depends on both the current CPU load and the host energy profile.
30 According to our measurements, the consumption is somehow linear in the amount of cores at full speed, with an
31 abnormality when all the cores are idle. The full details are in `our scientific paper
32 <https://hal.inria.fr/hal-01523608>`_ on that topic.
34 As a result, our energy model takes 4 parameters:
36 - ``Idle`` wattage (i.e., instantaneous consumption in Watt) when your host is up and running, but without anything to
38 - ``Epsilon`` wattage when all cores are at 0 or epsilon%, but not in Idle state.
39 - ``AllCores`` wattage when all cores of the host are at 100%.
40 - ``Off`` wattage when the host is turned off.
42 Here is an example of XML declaration:
46 <host id="HostA" speed="100.0Mf" core="4">
47 <prop id="wattage_per_state" value="100.0:120.0:200.0" />
48 <prop id="wattage_off" value="10" />
51 If only two values are given, ``Idle`` is used for the missing ``Epsilon`` value.
53 This example gives the following parameters: ``Off`` is 10 Watts; ``Idle`` is 100 Watts; ``Epsilon`` is 120 Watts and
54 ``AllCores`` is 200 Watts.
55 This is enough to compute the wattage as a function of the amount of loaded cores:
60 <tr><th>#Cores loaded</th><th>Wattage</th><th>Explanation</th></tr>
61 <tr><td>0 (idle)</td><td> 100 Watts </td><td> Idle value</td></tr>
62 <tr><td>0 (not idle)</td><td> 120 Watts</td><td> Epsilon 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>
72 <h4>How does DVFS interact with the host energy model?</h4>
74 If your host has several DVFS levels (several pstates), then you should give the energetic profile of each pstate level:
78 <host id="HostC" speed="100.0Mf,50.0Mf,20.0Mf" core="4">
79 <prop id="wattage_per_state"
80 value="95.0:120.0:200.0, 93.0:115.0:170.0, 90.0:110.0:150.0" />
81 <prop id="wattage_off" value="10" />
84 This encodes the following values:
89 <tr><th>pstate</th><th>Performance</th><th>Idle</th><th>Epsilon</th><th>AllCores</th></tr>
90 <tr><td>0</td><td>100 Mflop/s</td><td>95 Watts</td><td>120 Watts</td><td>200 Watts</td></tr>
91 <tr><td>1</td><td>50 Mflop/s</td><td>93 Watts</td><td>115 Watts</td><td>170 Watts</td></tr>
92 <tr><td>2</td><td>20 Mflop/s</td><td>90 Watts</td><td>110 Watts</td><td>150 Watts</td></tr>
95 To change the pstate of a given CPU, use the following functions:
96 :cpp:func:`MSG_host_get_nb_pstates()`, :cpp:func:`simgrid::s4u::Host::set_pstate()`,
97 :cpp:func:`MSG_host_get_power_peak_at()`.
101 <h4>How accurate are these models?</h4>
103 This model cannot be more accurate than your instantiation: with the default values, your result will not be accurate at
104 all. You can still get accurate energy prediction, provided that you carefully instantiate the model.
105 The first step is to ensure that your timing prediction match perfectly. But this is only the first step of the path,
106 and you really want to read `this paper <https://hal.inria.fr/hal-01523608>`_ to see all what you need to do
107 before you can get accurate energy predictions.
112 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_energy, surf, "Logging specific to the SURF energy plugin");
114 // Forwards declaration needed to make this function a friend (because friends have external linkage by default)
115 static void on_simulation_end();
127 PowerRange(double idle, double epsilon, double max) : idle_(idle), epsilon_(epsilon), max_(max), slope_(max-epsilon) {}
131 simgrid::s4u::Host* host_ = nullptr;
132 /*< List of (idle_power, epsilon_power, max_power) tuple corresponding to each cpu pstate */
133 std::vector<PowerRange> power_range_watts_list_;
134 bool has_pstate_power_values_ = false; /*< Whether power consumption values were provided for all pstates */
136 /* We need to keep track of what pstate has been used, as we will sometimes be notified only *after* a pstate has been
137 * used (but we need to update the energy consumption with the old pstate!)
140 const int pstate_off_ = -1;
141 double watts_off_ = 0.0; /*< Consumption when the machine is turned off (shutdown) */
142 double total_energy_ = 0.0; /*< Total energy consumed by the host */
143 double last_updated_ = surf_get_clock(); /*< Timestamp of the last energy update event*/
145 /* Only used to split total energy into unused/used hosts.
146 * If you want to get this info for something else, rather use the host_load plugin
148 bool host_was_used_ = false;
150 void init_watts_range_list();
151 friend void ::on_simulation_end(); // For access to host_was_used_
154 static simgrid::xbt::Extension<simgrid::s4u::Host, HostEnergy> EXTENSION_ID;
156 explicit HostEnergy(simgrid::s4u::Host* ptr);
159 bool has_pstate_power_values() const;
161 double get_current_watts_value();
162 double get_current_watts_value(double cpu_load) const;
163 double get_consumed_energy();
164 double get_watt_idle_at(int pstate) const;
165 double get_watt_min_at(int pstate) const;
166 double get_watt_max_at(int pstate) const;
167 double get_power_range_slope_at(int pstate) const;
168 double get_last_update_time() const { return last_updated_; }
172 simgrid::xbt::Extension<simgrid::s4u::Host, HostEnergy> HostEnergy::EXTENSION_ID;
174 /* Returns whether power consumption values were provided for all pstates. */
175 bool HostEnergy::has_pstate_power_values() const {
176 return has_pstate_power_values_;
179 /* Computes the consumption so far. Called lazily on need. */
180 void HostEnergy::update()
182 double start_time = last_updated_;
183 double finish_time = surf_get_clock();
185 // We may have start == finish if the past consumption was updated since the simcall was started
186 // for example if 2 actors requested to update the same host's consumption in a given scheduling round.
188 // Even in this case, we need to save the pstate for the next call (after this if),
189 // which may have changed since that recent update.
190 if (start_time < finish_time) {
191 double previous_energy = total_energy_;
193 double instantaneous_power_consumption = this->get_current_watts_value();
195 double energy_this_step = instantaneous_power_consumption * (finish_time - start_time);
197 // TODO Trace: Trace energy_this_step from start_time to finish_time in host->getName()
199 total_energy_ = previous_energy + energy_this_step;
200 last_updated_ = finish_time;
202 XBT_DEBUG("[update_energy of %s] period=[%.8f-%.8f]; current speed=%.2E flop/s (pstate %i); total consumption "
203 "before: %.8f J -> added now: %.8f J",
204 host_->get_cname(), start_time, finish_time, host_->get_pstate_speed(pstate_), pstate_, previous_energy,
208 /* Save data for the upcoming time interval: whether it's on/off and the pstate if it's on */
209 pstate_ = host_->is_on() ? host_->get_pstate() : pstate_off_;
212 HostEnergy::HostEnergy(simgrid::s4u::Host* ptr) : host_(ptr)
214 init_watts_range_list();
216 const char* off_power_str = host_->get_property("wattage_off");
217 if (off_power_str != nullptr) {
219 this->watts_off_ = std::stod(std::string(off_power_str));
220 } catch (const std::invalid_argument&) {
221 throw std::invalid_argument(std::string("Invalid value for property wattage_off of host ") + host_->get_cname() +
222 ": " + off_power_str);
225 /* watts_off is 0 by default */
228 HostEnergy::~HostEnergy() = default;
230 double HostEnergy::get_watt_idle_at(int pstate) const
232 if (not has_pstate_power_values_)
234 return power_range_watts_list_[pstate].idle_;
237 double HostEnergy::get_watt_min_at(int pstate) const
239 if (not has_pstate_power_values_)
241 return power_range_watts_list_[pstate].epsilon_;
244 double HostEnergy::get_watt_max_at(int pstate) const
246 if (not has_pstate_power_values_)
248 return power_range_watts_list_[pstate].max_;
251 double HostEnergy::get_power_range_slope_at(int pstate) const
253 if (not has_pstate_power_values_)
255 return power_range_watts_list_[pstate].slope_;
258 /** @brief Computes the power consumed by the host according to the current situation
260 * - If the host is off, that's the watts_off value
261 * - if it's on, take the current pstate and the current processor load into account */
262 double HostEnergy::get_current_watts_value()
264 if (this->pstate_ == pstate_off_) // The host is off (or was off at the beginning of this time interval)
265 return this->watts_off_;
267 double current_speed = host_->get_pstate_speed(this->pstate_);
271 if (current_speed <= 0)
272 // Some users declare a pstate of speed 0 flops (e.g., to model boot time).
273 // We consider that the machine is then fully loaded. That's arbitrary but it avoids a NaN
276 cpu_load = host_->get_load() / current_speed;
278 /* Divide by the number of cores here to have a value between 0 and 1 */
279 cpu_load /= host_->get_core_count();
281 if (cpu_load > 1) // This condition is true for energy_ptask on 32 bits, even if cpu_load is displayed as 1.000000
282 cpu_load = 1; // That may be an harmless rounding error?
284 host_was_used_ = true;
287 return get_current_watts_value(cpu_load);
290 /** @brief Computes the power that the host would consume at the provided processor load
292 * Whether the host is ON or OFF is not taken into account.
294 double HostEnergy::get_current_watts_value(double cpu_load) const
296 if (not has_pstate_power_values_)
299 /* Return watts_off if pstate == pstate_off (ie, if the host is off) */
300 if (this->pstate_ == pstate_off_) {
304 PowerRange power_range = power_range_watts_list_.at(this->pstate_);
305 double current_power;
310 * Something is going on, the host is not idle.
312 * The power consumption follows the regular model:
313 * P(cpu_load) = Pstatic + Pdynamic * cpu_load
314 * where Pstatic = power_range.epsilon_ and Pdynamic = power_range.slope_
315 * and the cpu_load is a value between 0 and 1.
317 current_power = power_range.epsilon_ + cpu_load * power_range.slope_;
321 /* The host is idle, take the dedicated value! */
322 current_power = power_range.idle_;
325 XBT_DEBUG("[get_current_watts] pstate=%i, epsilon_power=%f, max_power=%f, slope=%f", this->pstate_, power_range.epsilon_,
326 power_range.max_, power_range.slope_);
327 XBT_DEBUG("[get_current_watts] Current power (watts) = %f, load = %f", current_power, cpu_load);
329 return current_power;
332 double HostEnergy::get_consumed_energy()
334 if (last_updated_ < surf_get_clock()) // We need to simcall this as it modifies the environment
335 simgrid::kernel::actor::simcall(std::bind(&HostEnergy::update, this));
337 return total_energy_;
340 void HostEnergy::init_watts_range_list()
342 const char* all_power_values_str = host_->get_property("wattage_per_state");
343 if (all_power_values_str == nullptr) {
344 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());
348 std::vector<std::string> all_power_values;
349 boost::split(all_power_values, all_power_values_str, boost::is_any_of(","));
350 XBT_DEBUG("%s: power properties: %s", host_->get_cname(), all_power_values_str);
352 xbt_assert(all_power_values.size() == (unsigned)host_->get_pstate_count(),
353 "Invalid XML file. Found %zu energetic profiles for %d pstates", all_power_values.size(),
354 host_->get_pstate_count());
357 for (auto const& current_power_values_str : all_power_values) {
358 /* retrieve the power values associated with the pstate i */
359 std::vector<std::string> current_power_values;
360 boost::split(current_power_values, current_power_values_str, boost::is_any_of(":"));
362 xbt_assert(current_power_values.size() == 2 || current_power_values.size() == 3,
363 "Power properties incorrectly defined for host %s."
364 "It should be 'Idle:AllCores' (or 'Idle:Epsilon:AllCores') power values.",
368 double epsilon_power;
371 auto msg_idle = xbt::string_printf("Invalid Idle value for pstate %d on host %s", i, host_->get_cname());
372 auto msg_epsilon = xbt::string_printf("Invalid Epsilon value for pstate %d on host %s", i, host_->get_cname());
373 auto msg_max = xbt::string_printf("Invalid AllCores value for pstate %d on host %s", i, host_->get_cname());
375 idle_power = xbt_str_parse_double((current_power_values.at(0)).c_str(), msg_idle.c_str());
376 if (current_power_values.size() == 2) { // Case: Idle:AllCores
377 epsilon_power = xbt_str_parse_double((current_power_values.at(0)).c_str(), msg_idle.c_str());
378 max_power = xbt_str_parse_double((current_power_values.at(1)).c_str(), msg_max.c_str());
379 } else { // Case: Idle:Epsilon:AllCores
380 epsilon_power = xbt_str_parse_double((current_power_values.at(1)).c_str(), msg_epsilon.c_str());
381 max_power = xbt_str_parse_double((current_power_values.at(2)).c_str(), msg_max.c_str());
384 XBT_DEBUG("Creating PowerRange for host %s. Idle:%f, Epsilon:%f, AllCores:%f.", host_->get_cname(), idle_power, epsilon_power, max_power);
386 PowerRange range(idle_power, epsilon_power, max_power);
387 power_range_watts_list_.push_back(range);
391 has_pstate_power_values_ = true;
393 } // namespace plugin
394 } // namespace simgrid
396 using simgrid::plugin::HostEnergy;
398 /* **************************** events callback *************************** */
399 static void on_creation(simgrid::s4u::Host& host)
401 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
404 // TODO Trace: set to zero the energy variable associated to host->get_name()
406 host.extension_set(new HostEnergy(&host));
409 static void on_action_state_change(simgrid::kernel::resource::CpuAction const& action,
410 simgrid::kernel::resource::Action::State /*previous*/)
412 for (simgrid::kernel::resource::CpuImpl* const& cpu : action.cpus()) {
413 simgrid::s4u::Host* host = cpu->get_iface();
414 if (host != nullptr) {
415 // If it's a VM, take the corresponding PM
416 const simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
417 if (vm) // If it's a VM, take the corresponding PM
420 // Get the host_energy extension for the relevant host
421 auto* host_energy = host->extension<HostEnergy>();
423 if (host_energy->get_last_update_time() < surf_get_clock())
424 host_energy->update();
429 /* This callback is fired either when the host changes its state (on/off) ("onStateChange") or its speed
430 * (because the user changed the pstate, or because of external trace events) ("onSpeedChange") */
431 static void on_host_change(simgrid::s4u::Host const& host)
433 if (dynamic_cast<simgrid::s4u::VirtualMachine const*>(&host)) // Ignore virtual machines
436 auto* host_energy = host.extension<HostEnergy>();
438 host_energy->update();
441 static void on_host_destruction(simgrid::s4u::Host const& host)
443 if (dynamic_cast<simgrid::s4u::VirtualMachine const*>(&host)) // Ignore virtual machines
446 XBT_INFO("Energy consumption of host %s: %f Joules", host.get_cname(),
447 host.extension<HostEnergy>()->get_consumed_energy());
450 static void on_simulation_end()
452 double total_energy = 0.0; // Total energy consumption (whole platform)
453 double used_hosts_energy = 0.0; // Energy consumed by hosts that computed something
454 for (simgrid::s4u::Host const* host : simgrid::s4u::Engine::get_instance()->get_all_hosts()) {
455 if (host && dynamic_cast<const simgrid::s4u::VirtualMachine*>(host) == nullptr) { // Ignore virtual machines
456 double energy = host->extension<HostEnergy>()->get_consumed_energy();
457 total_energy += energy;
458 if (host->extension<HostEnergy>()->host_was_used_)
459 used_hosts_energy += energy;
462 XBT_INFO("Total energy consumption: %f Joules (used hosts: %f Joules; unused/idle hosts: %f)", total_energy,
463 used_hosts_energy, total_energy - used_hosts_energy);
466 /* **************************** Public interface *************************** */
468 /** @ingroup plugin_host_energy
469 * @brief Enable host energy plugin
470 * @details Enable energy plugin to get joules consumption of each cpu. Call this function before loading your platform.
472 void sg_host_energy_plugin_init()
474 if (HostEnergy::EXTENSION_ID.valid())
477 HostEnergy::EXTENSION_ID = simgrid::s4u::Host::extension_create<HostEnergy>();
479 simgrid::s4u::Host::on_creation.connect(&on_creation);
480 simgrid::s4u::Host::on_state_change.connect(&on_host_change);
481 simgrid::s4u::Host::on_speed_change.connect(&on_host_change);
482 simgrid::s4u::Host::on_destruction.connect(&on_host_destruction);
483 simgrid::s4u::Engine::on_simulation_end.connect(&on_simulation_end);
484 simgrid::kernel::resource::CpuAction::on_state_change.connect(&on_action_state_change);
485 // We may only have one actor on a node. If that actor executes something like
486 // compute -> recv -> compute
487 // the recv operation will not trigger a "CpuAction::on_state_change". This means
488 // that the next trigger would be the 2nd compute, hence ignoring the idle time
489 // during the recv call. By updating at the beginning of a compute, we can
490 // fix that. (If the cpu is not idle, this is not required.)
491 simgrid::s4u::Exec::on_start.connect([](simgrid::s4u::Exec const& activity) {
492 if (activity.get_host_number() == 1) { // We only run on one host
493 simgrid::s4u::Host* host = activity.get_host();
494 const simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
497 xbt_assert(host != nullptr);
498 host->extension<HostEnergy>()->update();
503 /** @ingroup plugin_host_energy
504 * @brief updates the consumption of all hosts
506 * After this call, sg_host_get_consumed_energy() will not interrupt your process
507 * (until after the next clock update).
509 void sg_host_energy_update_all()
511 simgrid::kernel::actor::simcall([]() {
512 std::vector<simgrid::s4u::Host*> list = simgrid::s4u::Engine::get_instance()->get_all_hosts();
513 for (auto const& host : list)
514 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(host) == nullptr) { // Ignore virtual machines
515 xbt_assert(host != nullptr);
516 host->extension<HostEnergy>()->update();
521 static void ensure_plugin_inited()
523 if (not HostEnergy::EXTENSION_ID.valid())
524 throw simgrid::xbt::InitializationError("The Energy plugin is not active. Please call sg_host_energy_plugin_init() "
525 "before calling any function related to that plugin.");
528 /** @ingroup plugin_host_energy
529 * @brief Returns the total energy consumed by the host so far (in Joules)
531 * Please note that since the consumption is lazily updated, it may require a simcall to update it.
532 * The result is that the actor requesting this value will be interrupted,
533 * the value will be updated in kernel mode before returning the control to the requesting actor.
535 double sg_host_get_consumed_energy(const_sg_host_t host)
537 ensure_plugin_inited();
538 auto host_energy = host->extension<HostEnergy>();
539 xbt_assert(host_energy->has_pstate_power_values(), "No power range properties specified for host %s",
541 return host_energy->get_consumed_energy();
544 /** @ingroup plugin_host_energy
545 * @brief Get the amount of watt dissipated when the host is idling
547 double sg_host_get_idle_consumption(const_sg_host_t host)
549 ensure_plugin_inited();
550 return host->extension<HostEnergy>()->get_watt_idle_at(0);
553 /** @ingroup plugin_host_energy
554 * @brief Get the amount of watt dissipated at the given pstate when the host is idling
556 double sg_host_get_idle_consumption_at(const_sg_host_t host, int pstate)
558 ensure_plugin_inited();
559 return host->extension<HostEnergy>()->get_watt_idle_at(pstate);
562 /** @ingroup plugin_host_energy
563 * @brief Get the amount of watt dissipated at the given pstate when the host is at 0 or epsilon% CPU usage.
565 double sg_host_get_wattmin_at(const_sg_host_t host, int pstate)
567 ensure_plugin_inited();
568 return host->extension<HostEnergy>()->get_watt_min_at(pstate);
570 /** @ingroup plugin_host_energy
571 * @brief Returns the amount of watt dissipated at the given pstate when the host burns CPU at 100%
573 double sg_host_get_wattmax_at(const_sg_host_t host, int pstate)
575 ensure_plugin_inited();
576 return host->extension<HostEnergy>()->get_watt_max_at(pstate);
578 /** @ingroup plugin_host_energy
579 * @brief Returns the power slope at the given pstate
581 double sg_host_get_power_range_slope_at(const_sg_host_t host, int pstate)
583 ensure_plugin_inited();
584 return host->extension<HostEnergy>()->get_power_range_slope_at(pstate);
586 /** @ingroup plugin_host_energy
587 * @brief Returns the current consumption of the host
589 double sg_host_get_current_consumption(const_sg_host_t host)
591 ensure_plugin_inited();
592 auto host_energy = host->extension<HostEnergy>();
593 xbt_assert(host_energy->has_pstate_power_values(), "No power range properties specified for host %s",
595 return host_energy->get_current_watts_value();