1 /* Copyright (c) 2010-2018. 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 OneCore: instantaneous consumption (in Watt) when only one core is active, at 100%.
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" power="100.0Mf" cores="4">
41 <prop id="watt_per_state" value="100.0:120.0:200.0" />
42 <prop id="watt_off" value="10" />
46 This example gives the following parameters: @b Off is 10 Watts; @b Idle is 100 Watts; @b OneCore is 120 Watts and @b
47 AllCores is 200 Watts.
48 This is enough to compute the consumption as a function of the amount of loaded cores:
51 <tr><th>@#Cores loaded</th><th>Consumption</th><th>Explanation</th></tr>
52 <tr><td>0</td><td> 100 Watts</td><td>Idle value</td></tr>
53 <tr><td>1</td><td> 120 Watts</td><td>OneCore value</td></tr>
54 <tr><td>2</td><td> 147 Watts</td><td>linear extrapolation between OneCore and AllCores</td></tr>
55 <tr><td>3</td><td> 173 Watts</td><td>linear extrapolation between OneCore and AllCores</td></tr>
56 <tr><td>4</td><td> 200 Watts</td><td>AllCores value</td></tr>
59 ### What if a given core is only at load 50%?
61 This is impossible in SimGrid because we recompute everything each time that the CPU starts or stops doing something.
62 So if a core is at load 50% over a period, it means that it is at load 100% half of the time and at load 0% the rest of
63 the time, and our model holds.
65 ### What if the host has only one core?
67 In this case, the parameters @b OneCore and @b AllCores are obviously the same.
68 Actually, SimGrid expect an energetic profile formatted as 'Idle:Running' for mono-cores hosts.
69 If you insist on passing 3 parameters in this case, then you must have the same value for @b OneCore and @b AllCores.
72 <host id="HostC" power="100.0Mf" cores="1">
73 <prop id="watt_per_state" value="95.0:200.0" /> <!-- we may have used '95:200:200' instead -->
74 <prop id="watt_off" value="10" />
78 ### How does DVFS interact with the host energy model?
80 If your host has several DVFS levels (several pstates), then you should give the energetic profile of each pstate level:
83 <host id="HostC" power="100.0Mf,50.0Mf,20.0Mf" cores="4">
84 <prop id="watt_per_state" value="95.0:120.0:200.0, 93.0:115.0:170.0, 90.0:110.0:150.0" />
85 <prop id="watt_off" value="10" />
89 This encodes the following values
91 <tr><th>pstate</th><th>Performance</th><th>Idle</th><th>OneCore</th><th>AllCores</th></tr>
92 <tr><td>0</td><td>100 Mflop/s</td><td>95 Watts</td><td>120 Watts</td><td>200 Watts</td></tr>
93 <tr><td>1</td><td>50 Mflop/s</td><td>93 Watts</td><td>115 Watts</td><td>170 Watts</td></tr>
94 <tr><td>2</td><td>20 Mflop/s</td><td>90 Watts</td><td>110 Watts</td><td>150 Watts</td></tr>
97 To change the pstate of a given CPU, use the following functions:
98 #MSG_host_get_nb_pstates(), simgrid#s4u#Host#setPstate(), #MSG_host_get_power_peak_at().
100 ### How accurate are these models?
102 This model cannot be more accurate than your instantiation: with the default values, your result will not be accurate at
103 all. You can still get accurate energy prediction, provided that you carefully instantiate the model.
104 The first step is to ensure that your timing prediction match perfectly. But this is only the first step of the path,
105 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
106 before you can get accurate energy predictions.
109 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_energy, surf, "Logging specific to the SURF energy plugin");
111 // Forwards declaration needed to make this function a friend (because friends have external linkage by default)
112 static void on_simulation_end();
123 PowerRange(double idle, double min, double max) : idle_(idle), min_(min), max_(max) {}
127 friend void ::on_simulation_end(); // For access to host_was_used_
129 static simgrid::xbt::Extension<simgrid::s4u::Host, HostEnergy> EXTENSION_ID;
131 explicit HostEnergy(simgrid::s4u::Host* ptr);
134 double get_current_watts_value();
135 double get_current_watts_value(double cpu_load);
136 double get_consumed_energy();
137 double get_watt_min_at(int pstate);
138 double get_watt_max_at(int pstate);
142 void init_watts_range_list();
143 simgrid::s4u::Host* host_ = nullptr;
144 /*< List of (min_power,max_power) pairs corresponding to each cpu pstate */
145 std::vector<PowerRange> power_range_watts_list_;
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;
153 /* Only used to split total energy into unused/used hosts.
154 * If you want to get this info for something else, rather use the host_load plugin
156 bool host_was_used_ = false;
158 double watts_off_ = 0.0; /*< Consumption when the machine is turned off (shutdown) */
159 double total_energy_ = 0.0; /*< Total energy consumed by the host */
160 double last_updated_; /*< Timestamp of the last energy update event*/
163 simgrid::xbt::Extension<simgrid::s4u::Host, HostEnergy> HostEnergy::EXTENSION_ID;
165 /* Computes the consumption so far. Called lazily on need. */
166 void HostEnergy::update()
168 double start_time = this->last_updated_;
169 double finish_time = surf_get_clock();
171 // We may have start == finish if the past consumption was updated since the simcall was started
172 // for example if 2 actors requested to update the same host's consumption in a given scheduling round.
174 // Even in this case, we need to save the pstate for the next call (after this if),
175 // which may have changed since that recent update.
176 if (start_time < finish_time) {
177 double previous_energy = this->total_energy_;
179 double instantaneous_consumption = this->get_current_watts_value();
181 double energy_this_step = instantaneous_consumption * (finish_time - start_time);
183 // TODO Trace: Trace energy_this_step from start_time to finish_time in host->getName()
185 this->total_energy_ = previous_energy + energy_this_step;
186 this->last_updated_ = finish_time;
188 XBT_DEBUG("[update_energy of %s] period=[%.8f-%.8f]; current speed=%.2E flop/s (pstate %i); total consumption before: consumption change: %.8f J -> added now: %.8f J",
189 host_->get_cname(), start_time, finish_time, host_->pimpl_cpu->get_pstate_peak_speed(this->pstate_), this->pstate_, previous_energy,
193 /* Save data for the upcoming time interval: whether it's on/off and the pstate if it's on */
194 this->pstate_ = host_->is_on() ? host_->get_pstate() : pstate_off_;
197 HostEnergy::HostEnergy(simgrid::s4u::Host* ptr) : host_(ptr), last_updated_(surf_get_clock())
199 init_watts_range_list();
201 const char* off_power_str = host_->get_property("watt_off");
202 if (off_power_str != nullptr) {
204 this->watts_off_ = std::stod(std::string(off_power_str));
205 } catch (std::invalid_argument& ia) {
206 throw std::invalid_argument(std::string("Invalid value for property watt_off of host ") + host_->get_cname() +
207 ": " + off_power_str);
210 /* watts_off is 0 by default */
213 HostEnergy::~HostEnergy() = default;
215 double HostEnergy::get_watt_min_at(int pstate)
217 xbt_assert(not power_range_watts_list_.empty(), "No power range properties specified for host %s",
219 return power_range_watts_list_[pstate].min_;
222 double HostEnergy::get_watt_max_at(int pstate)
224 xbt_assert(not power_range_watts_list_.empty(), "No power range properties specified for host %s",
226 return power_range_watts_list_[pstate].max_;
229 /** @brief Computes the power consumed by the host according to the current situation
231 * - If the host is off, that's the watts_off value
232 * - if it's on, take the current pstate and the current processor load into account */
233 double HostEnergy::get_current_watts_value()
235 if (this->pstate_ == pstate_off_) // The host is off (or was off at the beginning of this time interval)
236 return this->watts_off_;
238 double current_speed = host_->get_speed();
242 if (current_speed <= 0)
243 // Some users declare a pstate of speed 0 flops (e.g., to model boot time).
244 // We consider that the machine is then fully loaded. That's arbitrary but it avoids a NaN
247 cpu_load = host_->pimpl_cpu->get_constraint()->get_usage() / current_speed;
249 /** Divide by the number of cores here **/
250 cpu_load /= host_->pimpl_cpu->get_core_count();
252 if (cpu_load > 1) // A machine with a load > 1 consumes as much as a fully loaded machine, not more
255 host_was_used_ = true;
257 /* The problem with this model is that the load is always 0 or 1, never something less.
258 * Another possibility could be to model the total energy as
260 * X/(X+Y)*W_idle + Y/(X+Y)*W_burn
262 * where X is the amount of idling cores, and Y the amount of computing cores.
264 return get_current_watts_value(cpu_load);
267 /** @brief Computes the power that the host would consume at the provided processor load
269 * Whether the host is ON or OFF is not taken into account.
271 double HostEnergy::get_current_watts_value(double cpu_load)
273 xbt_assert(not power_range_watts_list_.empty(), "No power range properties specified for host %s",
276 /* Return watts_off if pstate == pstate_off (ie, if the host is off) */
277 if (this->pstate_ == pstate_off_) {
281 /* min_power corresponds to the power consumed when only one core is active */
282 /* max_power is the power consumed at 100% cpu load */
283 auto range = power_range_watts_list_.at(this->pstate_);
284 double current_power = 0;
285 double min_power = 0;
286 double max_power = 0;
287 double power_slope = 0;
289 if (cpu_load > 0) { /* Something is going on, the machine is not idle */
290 min_power = range.min_;
291 max_power = range.max_;
294 * The min_power states how much we consume when only one single
295 * core is working. This means that when cpu_load == 1/coreCount, then
296 * current_power == min_power.
298 * The maximum must be reached when all cores are working (but 1 core was
299 * already accounted for by min_power)
300 * i.e., we need min_power + (maxCpuLoad-1/coreCount)*power_slope == max_power
301 * (maxCpuLoad is by definition 1)
304 int coreCount = host_->get_core_count();
305 double coreReciprocal = static_cast<double>(1) / static_cast<double>(coreCount);
307 power_slope = (max_power - min_power) / (1 - coreReciprocal);
309 power_slope = 0; // Should be 0, since max_power == min_power (in this case)
311 current_power = min_power + (cpu_load - coreReciprocal) * power_slope;
312 } else { /* Our machine is idle, take the dedicated value! */
313 current_power = range.idle_;
316 XBT_DEBUG("[get_current_watts] pstate=%i, min_power=%f, max_power=%f, slope=%f", this->pstate_, min_power, max_power, power_slope);
317 XBT_DEBUG("[get_current_watts] Current power (watts) = %f, load = %f", current_power, cpu_load);
319 return current_power;
322 double HostEnergy::get_consumed_energy()
324 if (last_updated_ < surf_get_clock()) // We need to simcall this as it modifies the environment
325 simgrid::simix::simcall(std::bind(&HostEnergy::update, this));
327 return total_energy_;
330 void HostEnergy::init_watts_range_list()
332 const char* all_power_values_str = host_->get_property("watt_per_state");
333 if (all_power_values_str == nullptr)
336 std::vector<std::string> all_power_values;
337 boost::split(all_power_values, all_power_values_str, boost::is_any_of(","));
338 XBT_DEBUG("%s: profile: %s, cores: %d", host_->get_cname(), all_power_values_str, host_->get_core_count());
341 for (auto const& current_power_values_str : all_power_values) {
342 /* retrieve the power values associated with the current pstate */
343 std::vector<std::string> current_power_values;
344 boost::split(current_power_values, current_power_values_str, boost::is_any_of(":"));
345 if (host_->get_core_count() == 1) {
346 xbt_assert(current_power_values.size() == 2 || current_power_values.size() == 3,
347 "Power properties incorrectly defined for host %s."
348 "It should be 'Idle:FullSpeed' power values because you have one core only.",
350 if (current_power_values.size() == 2) {
351 // In this case, 1core == AllCores
352 current_power_values.push_back(current_power_values.at(1));
353 } else { // size == 3
354 current_power_values[2] = current_power_values.at(1);
355 static bool displayed_warning = false;
356 if (not displayed_warning) { // Otherwise we get in the worst case no_pstate*no_hosts warnings
357 XBT_WARN("Host %s is a single-core machine and part of the power profile is '%s'"
358 ", which is in the 'Idle:OneCore:AllCores' format."
359 " Since this is a single-core machine, AllCores and OneCore are identical."
360 " Here, only the value for 'OneCore' is used.", host_->get_cname(), current_power_values_str.c_str());
361 displayed_warning = true;
365 xbt_assert(current_power_values.size() == 3,
366 "Power properties incorrectly defined for host %s."
367 "It should be 'Idle:OneCore:AllCores' power values because you have more than one core.",
371 /* min_power corresponds to the idle power (cpu load = 0) */
372 /* max_power is the power consumed at 100% cpu load */
373 char* msg_idle = bprintf("Invalid idle value for pstate %d on host %s: %%s", i, host_->get_cname());
374 char* msg_min = bprintf("Invalid OneCore value for pstate %d on host %s: %%s", i, host_->get_cname());
375 char* msg_max = bprintf("Invalid AllCores value for pstate %d on host %s: %%s", i, host_->get_cname());
376 PowerRange range(xbt_str_parse_double((current_power_values.at(0)).c_str(), msg_idle),
377 xbt_str_parse_double((current_power_values.at(1)).c_str(), msg_min),
378 xbt_str_parse_double((current_power_values.at(2)).c_str(), msg_max));
379 power_range_watts_list_.push_back(range);
386 } // namespace plugin
387 } // namespace simgrid
389 using simgrid::plugin::HostEnergy;
391 /* **************************** events callback *************************** */
392 static void on_creation(simgrid::s4u::Host& host)
394 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
397 // TODO Trace: set to zero the energy variable associated to host->getName()
399 host.extension_set(new HostEnergy(&host));
402 static void on_action_state_change(simgrid::surf::CpuAction* action,
403 simgrid::kernel::resource::Action::State /*previous*/)
405 for (simgrid::surf::Cpu* const& cpu : action->cpus()) {
406 simgrid::s4u::Host* host = cpu->get_host();
407 if (host != nullptr) {
409 // If it's a VM, take the corresponding PM
410 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
411 if (vm) // If it's a VM, take the corresponding PM
414 // Get the host_energy extension for the relevant host
415 HostEnergy* host_energy = host->extension<HostEnergy>();
417 if (host_energy->last_updated_ < surf_get_clock())
418 host_energy->update();
423 /* This callback is fired either when the host changes its state (on/off) ("onStateChange") or its speed
424 * (because the user changed the pstate, or because of external trace events) ("onSpeedChange") */
425 static void on_host_change(simgrid::s4u::Host& host)
427 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
430 HostEnergy* host_energy = host.extension<HostEnergy>();
432 host_energy->update();
435 static void on_host_destruction(simgrid::s4u::Host& host)
437 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
440 XBT_INFO("Energy consumption of host %s: %f Joules", host.get_cname(),
441 host.extension<HostEnergy>()->get_consumed_energy());
444 static void on_simulation_end()
446 std::vector<simgrid::s4u::Host*> hosts = simgrid::s4u::Engine::get_instance()->get_all_hosts();
448 double total_energy = 0.0; // Total energy consumption (whole platform)
449 double used_hosts_energy = 0.0; // Energy consumed by hosts that computed something
450 for (size_t i = 0; i < hosts.size(); i++) {
451 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(hosts[i]) == nullptr) { // Ignore virtual machines
453 double energy = hosts[i]->extension<HostEnergy>()->get_consumed_energy();
454 total_energy += energy;
455 if (hosts[i]->extension<HostEnergy>()->host_was_used_)
456 used_hosts_energy += energy;
459 XBT_INFO("Total energy consumption: %f Joules (used hosts: %f Joules; unused/idle hosts: %f)", total_energy,
460 used_hosts_energy, total_energy - used_hosts_energy);
463 /* **************************** Public interface *************************** */
465 /** @ingroup plugin_energy
466 * @brief Enable host energy plugin
467 * @details Enable energy plugin to get joules consumption of each cpu. Call this function before #MSG_init().
469 void sg_host_energy_plugin_init()
471 if (HostEnergy::EXTENSION_ID.valid())
474 HostEnergy::EXTENSION_ID = simgrid::s4u::Host::extension_create<HostEnergy>();
476 simgrid::s4u::Host::on_creation.connect(&on_creation);
477 simgrid::s4u::Host::on_state_change.connect(&on_host_change);
478 simgrid::s4u::Host::on_speed_change.connect(&on_host_change);
479 simgrid::s4u::Host::on_destruction.connect(&on_host_destruction);
480 simgrid::s4u::on_simulation_end.connect(&on_simulation_end);
481 simgrid::surf::CpuAction::on_state_change.connect(&on_action_state_change);
482 // We may only have one actor on a node. If that actor executes something like
483 // compute -> recv -> compute
484 // the recv operation will not trigger a "CpuAction::on_state_change". This means
485 // that the next trigger would be the 2nd compute, hence ignoring the idle time
486 // during the recv call. By updating at the beginning of a compute, we can
487 // fix that. (If the cpu is not idle, this is not required.)
488 simgrid::kernel::activity::ExecImpl::on_creation.connect([](simgrid::kernel::activity::ExecImplPtr activity){
489 if (activity->host_ != nullptr) { // We only run on one host
490 simgrid::s4u::Host* host = activity->host_;
491 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(activity->host_))
492 host = dynamic_cast<simgrid::s4u::VirtualMachine*>(activity->host_)->get_pm();
494 host->extension<HostEnergy>()->update();
499 /** @ingroup plugin_energy
500 * @brief updates the consumption of all hosts
502 * After this call, sg_host_get_consumed_energy() will not interrupt your process
503 * (until after the next clock update).
505 void sg_host_energy_update_all()
507 simgrid::simix::simcall([]() {
508 std::vector<simgrid::s4u::Host*> list = simgrid::s4u::Engine::get_instance()->get_all_hosts();
509 for (auto const& host : list)
510 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(host) == nullptr) // Ignore virtual machines
511 host->extension<HostEnergy>()->update();
515 /** @ingroup plugin_energy
516 * @brief Returns the total energy consumed by the host so far (in Joules)
518 * Please note that since the consumption is lazily updated, it may require a simcall to update it.
519 * The result is that the actor requesting this value will be interrupted,
520 * the value will be updated in kernel mode before returning the control to the requesting actor.
522 double sg_host_get_consumed_energy(sg_host_t host)
524 xbt_assert(HostEnergy::EXTENSION_ID.valid(),
525 "The Energy plugin is not active. Please call sg_host_energy_plugin_init() during initialization.");
526 return host->extension<HostEnergy>()->get_consumed_energy();
529 /** @ingroup plugin_energy
530 * @brief Get the amount of watt dissipated at the given pstate when the host is idling
532 double sg_host_get_wattmin_at(sg_host_t host, int pstate)
534 xbt_assert(HostEnergy::EXTENSION_ID.valid(),
535 "The Energy plugin is not active. Please call sg_host_energy_plugin_init() during initialization.");
536 return host->extension<HostEnergy>()->get_watt_min_at(pstate);
538 /** @ingroup plugin_energy
539 * @brief Returns the amount of watt dissipated at the given pstate when the host burns CPU at 100%
541 double sg_host_get_wattmax_at(sg_host_t host, int pstate)
543 xbt_assert(HostEnergy::EXTENSION_ID.valid(),
544 "The Energy plugin is not active. Please call sg_host_energy_plugin_init() during initialization.");
545 return host->extension<HostEnergy>()->get_watt_max_at(pstate);
548 /** @ingroup plugin_energy
549 * @brief Returns the current consumption of the host
551 double sg_host_get_current_consumption(sg_host_t host)
553 xbt_assert(HostEnergy::EXTENSION_ID.valid(),
554 "The Energy plugin is not active. Please call sg_host_energy_plugin_init() during initialization.");
555 return host->extension<HostEnergy>()->get_current_watts_value();