1 /* Copyright (c) 2010-2019. 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/plugins/energy.h"
7 #include "simgrid/s4u/Engine.hpp"
8 #include "src/kernel/activity/ExecImpl.hpp"
9 #include "src/include/surf/surf.hpp"
10 #include "src/plugins/vm/VirtualMachineImpl.hpp"
11 #include "src/surf/cpu_interface.hpp"
13 #include <boost/algorithm/string/classification.hpp>
14 #include <boost/algorithm/string/split.hpp>
16 SIMGRID_REGISTER_PLUGIN(host_energy, "Cpu energy consumption.", &sg_host_energy_plugin_init)
18 /** @addtogroup plugin_energy
20 This is the energy plugin, enabling to account not only for computation time, but also for the dissipated energy in the
22 To activate this plugin, first call sg_host_energy_plugin_init() before your #MSG_init(), and then use
23 MSG_host_get_consumed_energy() to retrieve the consumption of a given host.
25 When the host is on, this energy consumption naturally depends on both the current CPU load and the host energy profile.
26 According to our measurements, the consumption is somehow linear in the amount of cores at full speed, with an
27 abnormality when all the cores are idle. The full details are in
28 <a href="https://hal.inria.fr/hal-01523608">our scientific paper</a> on that topic.
30 As a result, our energy model takes 4 parameters:
32 - @b Idle: instantaneous consumption (in Watt) when your host is up and running, but without anything to do.
33 - @b Epsilon: instantaneous consumption (in Watt) when all cores are at 0 or epsilon%, but not in Idle state.
34 - @b AllCores: instantaneous consumption (in Watt) when all cores of the host are at 100%.
35 - @b Off: instantaneous consumption (in Watt) when the host is turned off.
37 Here is an example of XML declaration:
40 <host id="HostA" speed="100.0Mf" core="4">
41 <prop id="watt_per_state" value="100.0:120.0:200.0" />
42 <prop id="watt_off" value="10" />
46 Please note that the 'Epsilon' parameter can be omitted in the XML declaration. In that case, the value of 'Epsilon' will
47 be the same as 'Idle'.
50 This example gives the following parameters: @b Off is 10 Watts; @b Idle is 100 Watts; @b Epsilon is 120 Watts and @b
51 AllCores is 200 Watts.
52 This is enough to compute the consumption as a function of the amount of loaded cores:
55 <tr><th>@#Cores loaded</th><th>Consumption</th><th>Explanation</th></tr>
56 <tr><td>0 (idle)</td><td> 100 Watts</td><td>Idle value</td></tr>
57 <tr><td>0 (not idle)</td><td> 120 Watts</td><td>Epsilon value</td></tr>
58 <tr><td>1</td><td> 140 Watts</td><td>linear extrapolation between Epsilon and AllCores</td></tr>
59 <tr><td>2</td><td> 160 Watts</td><td>linear extrapolation between Epsilon and AllCores</td></tr>
60 <tr><td>3</td><td> 180 Watts</td><td>linear extrapolation between Epsilon and AllCores</td></tr>
61 <tr><td>4</td><td> 200 Watts</td><td>AllCores value</td></tr>
65 ### How does DVFS interact with the host energy model?
67 If your host has several DVFS levels (several pstates), then you should give the energetic profile of each pstate level:
70 <host id="HostC" speed="100.0Mf,50.0Mf,20.0Mf" core="4">
71 <prop id="watt_per_state" value="95.0:120.0:200.0, 93.0:115.0:170.0, 90.0:110.0:150.0" />
72 <prop id="watt_off" value="10" />
76 This encodes the following values
78 <tr><th>pstate</th><th>Performance</th><th>Idle</th><th>Epsilon</th><th>AllCores</th></tr>
79 <tr><td>0</td><td>100 Mflop/s</td><td>95 Watts</td><td>120 Watts</td><td>200 Watts</td></tr>
80 <tr><td>1</td><td>50 Mflop/s</td><td>93 Watts</td><td>115 Watts</td><td>170 Watts</td></tr>
81 <tr><td>2</td><td>20 Mflop/s</td><td>90 Watts</td><td>110 Watts</td><td>150 Watts</td></tr>
84 To change the pstate of a given CPU, use the following functions:
85 #MSG_host_get_nb_pstates(), simgrid#s4u#Host#setPstate(), #MSG_host_get_power_peak_at().
87 ### How accurate are these models?
89 This model cannot be more accurate than your instantiation: with the default values, your result will not be accurate at
90 all. You can still get accurate energy prediction, provided that you carefully instantiate the model.
91 The first step is to ensure that your timing prediction match perfectly. But this is only the first step of the path,
92 and you really want to read <a href="https://hal.inria.fr/hal-01523608">this paper</a> to see all what you need to do
93 before you can get accurate energy predictions.
96 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_energy, surf, "Logging specific to the SURF energy plugin");
98 // Forwards declaration needed to make this function a friend (because friends have external linkage by default)
99 static void on_simulation_end();
111 PowerRange(double idle, double epsilon, double max) : idle_(idle), epsilon_(epsilon), max_(max), slope_(max-epsilon) {}
115 friend void ::on_simulation_end(); // For access to host_was_used_
117 static simgrid::xbt::Extension<simgrid::s4u::Host, HostEnergy> EXTENSION_ID;
119 explicit HostEnergy(simgrid::s4u::Host* ptr);
122 double get_current_watts_value();
123 double get_current_watts_value(double cpu_load);
124 double get_consumed_energy();
125 double get_idle_consumption();
126 double get_watt_min_at(int pstate);
127 double get_watt_max_at(int pstate);
128 double get_power_range_slope_at(int pstate);
132 void init_watts_range_list();
133 simgrid::s4u::Host* host_ = nullptr;
134 /*< List of (idle_power, epsilon_power, max_power) tuple corresponding to each cpu pstate */
135 std::vector<PowerRange> power_range_watts_list_;
137 /* We need to keep track of what pstate has been used, as we will sometimes be notified only *after* a pstate has been
138 * used (but we need to update the energy consumption with the old pstate!)
141 const int pstate_off_ = -1;
143 /* Only used to split total energy into unused/used hosts.
144 * If you want to get this info for something else, rather use the host_load plugin
146 bool host_was_used_ = false;
148 double watts_off_ = 0.0; /*< Consumption when the machine is turned off (shutdown) */
149 double total_energy_ = 0.0; /*< Total energy consumed by the host */
150 double last_updated_; /*< Timestamp of the last energy update event*/
153 simgrid::xbt::Extension<simgrid::s4u::Host, HostEnergy> HostEnergy::EXTENSION_ID;
155 /* Computes the consumption so far. Called lazily on need. */
156 void HostEnergy::update()
158 double start_time = this->last_updated_;
159 double finish_time = surf_get_clock();
161 // We may have start == finish if the past consumption was updated since the simcall was started
162 // for example if 2 actors requested to update the same host's consumption in a given scheduling round.
164 // Even in this case, we need to save the pstate for the next call (after this if),
165 // which may have changed since that recent update.
166 if (start_time < finish_time) {
167 double previous_energy = this->total_energy_;
169 double instantaneous_power_consumption = this->get_current_watts_value();
171 double energy_this_step = instantaneous_power_consumption * (finish_time - start_time);
173 // TODO Trace: Trace energy_this_step from start_time to finish_time in host->getName()
175 this->total_energy_ = previous_energy + energy_this_step;
176 this->last_updated_ = finish_time;
178 XBT_DEBUG("[update_energy of %s] period=[%.8f-%.8f]; current speed=%.2E flop/s (pstate %i); total consumption before: %.8f J -> added now: %.8f J",
179 host_->get_cname(), start_time, finish_time, host_->pimpl_cpu->get_pstate_peak_speed(this->pstate_), this->pstate_, previous_energy,
183 /* Save data for the upcoming time interval: whether it's on/off and the pstate if it's on */
184 this->pstate_ = host_->is_on() ? host_->get_pstate() : pstate_off_;
187 HostEnergy::HostEnergy(simgrid::s4u::Host* ptr) : host_(ptr), last_updated_(surf_get_clock())
189 init_watts_range_list();
191 const char* off_power_str = host_->get_property("watt_off");
192 if (off_power_str != nullptr) {
194 this->watts_off_ = std::stod(std::string(off_power_str));
195 } catch (const std::invalid_argument&) {
196 throw std::invalid_argument(std::string("Invalid value for property watt_off of host ") + host_->get_cname() +
197 ": " + off_power_str);
200 /* watts_off is 0 by default */
203 HostEnergy::~HostEnergy() = default;
205 double HostEnergy::get_idle_consumption()
207 xbt_assert(not power_range_watts_list_.empty(), "No power range properties specified for host %s",
210 return power_range_watts_list_[0].idle_;
213 double HostEnergy::get_watt_min_at(int pstate)
215 xbt_assert(not power_range_watts_list_.empty(), "No power range properties specified for host %s",
217 return power_range_watts_list_[pstate].epsilon_;
220 double HostEnergy::get_watt_max_at(int pstate)
222 xbt_assert(not power_range_watts_list_.empty(), "No power range properties specified for host %s",
224 return power_range_watts_list_[pstate].max_;
227 double HostEnergy::get_power_range_slope_at(int pstate)
229 xbt_assert(not power_range_watts_list_.empty(), "No power range properties specified for host %s",
231 return power_range_watts_list_[pstate].slope_;
234 /** @brief Computes the power consumed by the host according to the current situation
236 * - If the host is off, that's the watts_off value
237 * - if it's on, take the current pstate and the current processor load into account */
238 double HostEnergy::get_current_watts_value()
240 if (this->pstate_ == pstate_off_) // The host is off (or was off at the beginning of this time interval)
241 return this->watts_off_;
243 double current_speed = host_->get_pstate_speed(this->pstate_);
247 if (current_speed <= 0)
248 // Some users declare a pstate of speed 0 flops (e.g., to model boot time).
249 // We consider that the machine is then fully loaded. That's arbitrary but it avoids a NaN
252 cpu_load = host_->pimpl_cpu->get_constraint()->get_usage() / current_speed;
254 /** Divide by the number of cores here to have a value between 0 and 1 **/
255 cpu_load /= host_->pimpl_cpu->get_core_count();
257 if (cpu_load > 1) // A machine with a load > 1 consumes as much as a fully loaded machine, not more
260 host_was_used_ = true;
263 /* @mquinson: The problem with this model is that the load is always 0 or 1, never something less.
264 * Another possibility could be to model the total energy as
266 * X/(X+Y)*W_idle + Y/(X+Y)*W_burn
268 * where X is the amount of idling cores, and Y the amount of computing cores.
270 * @Mommessc: I do not think the load is always 0 or 1 anymore.
271 * Moreover, it is not quite clear how the regular model of power consumption (P = Pstatic + load * Pdynamic)
272 * is impacted if we separate the number of idle and working cores.
274 return get_current_watts_value(cpu_load);
277 /** @brief Computes the power that the host would consume at the provided processor load
279 * Whether the host is ON or OFF is not taken into account.
281 double HostEnergy::get_current_watts_value(double cpu_load)
283 xbt_assert(not power_range_watts_list_.empty(), "No power range properties specified for host %s",
286 /* Return watts_off if pstate == pstate_off (ie, if the host is off) */
287 if (this->pstate_ == pstate_off_) {
291 PowerRange power_range = power_range_watts_list_.at(this->pstate_);
292 double current_power;
297 * Something is going on, the host is not idle.
299 * The power consumption follows the regular model:
300 * P(cpu_load) = Pstatic + Pdynamic * cpu_load
301 * where Pstatic = power_range.epsilon_ and Pdynamic = power_range.slope_
302 * and the cpu_load is a value between 0 and 1.
304 current_power = power_range.epsilon_ + cpu_load * power_range.slope_;
308 /* The host is idle, take the dedicated value! */
309 current_power = power_range.idle_;
312 XBT_DEBUG("[get_current_watts] pstate=%i, epsilon_power=%f, max_power=%f, slope=%f", this->pstate_, power_range.epsilon_,
313 power_range.max_, power_range.slope_);
314 XBT_DEBUG("[get_current_watts] Current power (watts) = %f, load = %f", current_power, cpu_load);
316 return current_power;
319 double HostEnergy::get_consumed_energy()
321 if (last_updated_ < surf_get_clock()) // We need to simcall this as it modifies the environment
322 simgrid::kernel::actor::simcall(std::bind(&HostEnergy::update, this));
324 return total_energy_;
327 void HostEnergy::init_watts_range_list()
329 const char* all_power_values_str = host_->get_property("watt_per_state");
330 if (all_power_values_str == nullptr)
333 std::vector<std::string> all_power_values;
334 boost::split(all_power_values, all_power_values_str, boost::is_any_of(","));
335 XBT_DEBUG("%s: power properties: %s", host_->get_cname(), all_power_values_str);
338 for (auto const& current_power_values_str : all_power_values) {
339 /* retrieve the power values associated with the pstate i */
340 std::vector<std::string> current_power_values;
341 boost::split(current_power_values, current_power_values_str, boost::is_any_of(":"));
343 xbt_assert(current_power_values.size() == 2 || current_power_values.size() == 3,
344 "Power properties incorrectly defined for host %s."
345 "It should be 'Idle:AllCores' (or 'Idle:Epsilon:AllCores') power values.",
349 double epsilon_power;
352 char* msg_idle = bprintf("Invalid Idle value for pstate %d on host %s: %%s", i, host_->get_cname());
353 char* msg_epsilon = bprintf("Invalid Epsilon value for pstate %d on host %s: %%s", i, host_->get_cname());
354 char* msg_max = bprintf("Invalid AllCores value for pstate %d on host %s: %%s", i, host_->get_cname());
356 idle_power = xbt_str_parse_double((current_power_values.at(0)).c_str(), msg_idle);
357 if (current_power_values.size() == 2) // Case: Idle:AllCores
359 epsilon_power = xbt_str_parse_double((current_power_values.at(0)).c_str(), msg_idle);
360 max_power = xbt_str_parse_double((current_power_values.at(1)).c_str(), msg_max);
362 else // Case: Idle:Epsilon:AllCores
364 epsilon_power = xbt_str_parse_double((current_power_values.at(1)).c_str(), msg_epsilon);
365 max_power = xbt_str_parse_double((current_power_values.at(2)).c_str(), msg_max);
368 XBT_DEBUG("Creating PowerRange for host %s. Idle:%f, Epsilon:%f, AllCores:%f.", host_->get_cname(), idle_power, epsilon_power, max_power);
370 PowerRange range(idle_power, epsilon_power, max_power);
371 power_range_watts_list_.push_back(range);
373 xbt_free(msg_epsilon);
378 } // namespace plugin
379 } // namespace simgrid
381 using simgrid::plugin::HostEnergy;
383 /* **************************** events callback *************************** */
384 static void on_creation(simgrid::s4u::Host& host)
386 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
389 // TODO Trace: set to zero the energy variable associated to host->getName()
391 host.extension_set(new HostEnergy(&host));
394 static void on_action_state_change(simgrid::kernel::resource::CpuAction const& action,
395 simgrid::kernel::resource::Action::State /*previous*/)
397 for (simgrid::kernel::resource::Cpu* const& cpu : action.cpus()) {
398 simgrid::s4u::Host* host = cpu->get_host();
399 if (host != nullptr) {
401 // If it's a VM, take the corresponding PM
402 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
403 if (vm) // If it's a VM, take the corresponding PM
406 // Get the host_energy extension for the relevant host
407 HostEnergy* host_energy = host->extension<HostEnergy>();
409 if (host_energy->last_updated_ < surf_get_clock())
410 host_energy->update();
415 /* This callback is fired either when the host changes its state (on/off) ("onStateChange") or its speed
416 * (because the user changed the pstate, or because of external trace events) ("onSpeedChange") */
417 static void on_host_change(simgrid::s4u::Host const& host)
419 if (dynamic_cast<simgrid::s4u::VirtualMachine const*>(&host)) // Ignore virtual machines
422 HostEnergy* host_energy = host.extension<HostEnergy>();
424 host_energy->update();
427 static void on_host_destruction(simgrid::s4u::Host const& host)
429 if (dynamic_cast<simgrid::s4u::VirtualMachine const*>(&host)) // Ignore virtual machines
432 XBT_INFO("Energy consumption of host %s: %f Joules", host.get_cname(),
433 host.extension<HostEnergy>()->get_consumed_energy());
436 static void on_simulation_end()
438 std::vector<simgrid::s4u::Host*> hosts = simgrid::s4u::Engine::get_instance()->get_all_hosts();
440 double total_energy = 0.0; // Total energy consumption (whole platform)
441 double used_hosts_energy = 0.0; // Energy consumed by hosts that computed something
442 for (size_t i = 0; i < hosts.size(); i++) {
443 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(hosts[i]) == nullptr) { // Ignore virtual machines
445 double energy = hosts[i]->extension<HostEnergy>()->get_consumed_energy();
446 total_energy += energy;
447 if (hosts[i]->extension<HostEnergy>()->host_was_used_)
448 used_hosts_energy += energy;
451 XBT_INFO("Total energy consumption: %f Joules (used hosts: %f Joules; unused/idle hosts: %f)", total_energy,
452 used_hosts_energy, total_energy - used_hosts_energy);
455 /* **************************** Public interface *************************** */
457 /** @ingroup plugin_energy
458 * @brief Enable host energy plugin
459 * @details Enable energy plugin to get joules consumption of each cpu. Call this function before #MSG_init().
461 void sg_host_energy_plugin_init()
463 if (HostEnergy::EXTENSION_ID.valid())
466 HostEnergy::EXTENSION_ID = simgrid::s4u::Host::extension_create<HostEnergy>();
468 simgrid::s4u::Host::on_creation.connect(&on_creation);
469 simgrid::s4u::Host::on_state_change.connect(&on_host_change);
470 simgrid::s4u::Host::on_speed_change.connect(&on_host_change);
471 simgrid::s4u::Host::on_destruction.connect(&on_host_destruction);
472 simgrid::s4u::Engine::on_simulation_end.connect(&on_simulation_end);
473 simgrid::kernel::resource::CpuAction::on_state_change.connect(&on_action_state_change);
474 // We may only have one actor on a node. If that actor executes something like
475 // compute -> recv -> compute
476 // the recv operation will not trigger a "CpuAction::on_state_change". This means
477 // that the next trigger would be the 2nd compute, hence ignoring the idle time
478 // during the recv call. By updating at the beginning of a compute, we can
479 // fix that. (If the cpu is not idle, this is not required.)
480 simgrid::kernel::activity::ExecImpl::on_creation.connect([](simgrid::kernel::activity::ExecImpl const& activity) {
481 if (activity.get_host_number() == 1) { // We only run on one host
482 simgrid::s4u::Host* host = activity.get_host();
483 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
486 xbt_assert(host != nullptr);
487 host->extension<HostEnergy>()->update();
492 /** @ingroup plugin_energy
493 * @brief updates the consumption of all hosts
495 * After this call, sg_host_get_consumed_energy() will not interrupt your process
496 * (until after the next clock update).
498 void sg_host_energy_update_all()
500 simgrid::kernel::actor::simcall([]() {
501 std::vector<simgrid::s4u::Host*> list = simgrid::s4u::Engine::get_instance()->get_all_hosts();
502 for (auto const& host : list)
503 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(host) == nullptr) { // Ignore virtual machines
504 xbt_assert(host != nullptr);
505 host->extension<HostEnergy>()->update();
510 /** @ingroup plugin_energy
511 * @brief Returns the total energy consumed by the host so far (in Joules)
513 * Please note that since the consumption is lazily updated, it may require a simcall to update it.
514 * The result is that the actor requesting this value will be interrupted,
515 * the value will be updated in kernel mode before returning the control to the requesting actor.
517 double sg_host_get_consumed_energy(sg_host_t host)
519 xbt_assert(HostEnergy::EXTENSION_ID.valid(),
520 "The Energy plugin is not active. Please call sg_host_energy_plugin_init() during initialization.");
521 return host->extension<HostEnergy>()->get_consumed_energy();
524 /** @ingroup plugin_energy
525 * @brief Get the amount of watt dissipated when the host is idling
527 double sg_host_get_idle_consumption(sg_host_t host)
529 xbt_assert(HostEnergy::EXTENSION_ID.valid(),
530 "The Energy plugin is not active. Please call sg_host_energy_plugin_init() during initialization.");
531 return host->extension<HostEnergy>()->get_idle_consumption();
534 /** @ingroup plugin_energy
535 * @brief Get the amount of watt dissipated at the given pstate when the host is idling
537 double sg_host_get_wattmin_at(sg_host_t host, int pstate)
539 xbt_assert(HostEnergy::EXTENSION_ID.valid(),
540 "The Energy plugin is not active. Please call sg_host_energy_plugin_init() during initialization.");
541 return host->extension<HostEnergy>()->get_watt_min_at(pstate);
543 /** @ingroup plugin_energy
544 * @brief Returns the amount of watt dissipated at the given pstate when the host burns CPU at 100%
546 double sg_host_get_wattmax_at(sg_host_t host, int pstate)
548 xbt_assert(HostEnergy::EXTENSION_ID.valid(),
549 "The Energy plugin is not active. Please call sg_host_energy_plugin_init() during initialization.");
550 return host->extension<HostEnergy>()->get_watt_max_at(pstate);
552 /** @ingroup plugin_energy
553 * @brief Returns the power slope at the given pstate
555 double sg_host_get_power_range_slope_at(sg_host_t host, int pstate)
557 xbt_assert(HostEnergy::EXTENSION_ID.valid(),
558 "The Energy plugin is not active. Please call sg_host_energy_plugin_init() during initialization.");
559 return host->extension<HostEnergy>()->get_power_range_slope_at(pstate);
561 /** @ingroup plugin_energy
562 * @brief Returns the current consumption of the host
564 double sg_host_get_current_consumption(sg_host_t host)
566 xbt_assert(HostEnergy::EXTENSION_ID.valid(),
567 "The Energy plugin is not active. Please call sg_host_energy_plugin_init() during initialization.");
568 return host->extension<HostEnergy>()->get_current_watts_value();