1 /* Copyright (c) 2013-2023. 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. */
7 #include "simgrid/kernel/routing/NetZoneImpl.hpp"
8 #include "simgrid/s4u/Engine.hpp"
9 #include "src/kernel/EngineImpl.hpp"
10 #include "src/kernel/resource/profile/Event.hpp"
11 #include "src/kernel/resource/profile/Profile.hpp"
12 #include "xbt/asserts.h"
17 constexpr double EPSILON = 0.000000001;
19 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(cpu_ti, res_cpu, "CPU resource, Trace Integration model");
21 namespace simgrid::kernel::resource {
27 CpuTiProfile::CpuTiProfile(const profile::Profile* profile)
31 double prev_value = 1;
32 const std::vector<profile::DatedValue>& events = profile->get_event_list();
33 xbt_assert(not events.empty());
34 unsigned long nb_points = events.size() + 1;
35 time_points_.reserve(nb_points);
36 integral_.reserve(nb_points);
37 for (auto const& val : events) {
39 integral += val.date_ * prev_value;
40 time_points_.push_back(time);
41 integral_.push_back(integral);
42 prev_value = val.value_;
45 double delay = profile->get_repeat_delay() + events.at(0).date_;
47 xbt_assert(events.back().value_ == prev_value, "Profiles need to end as they start");
49 integral += delay * prev_value;
51 time_points_.push_back(time);
52 integral_.push_back(integral);
56 * @brief Integrate trace
58 * Wrapper around profile_->integrate_simple() to get
61 * @param a Begin of interval
62 * @param b End of interval
63 * @return the integrate value. -1 if an error occurs.
65 double CpuTiTmgr::integrate(double a, double b) const
67 xbt_assert(a >= 0.0 && a <= b,
68 "Error, invalid integration interval [%.2f,%.2f]. You probably have a task executing with negative "
69 "computation amount. Check your code.",
71 if (fabs(a - b) < EPSILON)
74 if (type_ == Type::FIXED) {
75 return (b - a) * value_;
79 if (fabs(ceil(a / last_time_) - a / last_time_) < EPSILON)
80 a_index = 1 + ceil(a / last_time_);
82 a_index = ceil(a / last_time_);
83 double b_index = floor(b / last_time_);
85 if (a_index > b_index) { /* Same chunk */
86 return profile_->integrate_simple(a - (a_index - 1) * last_time_, b - b_index * last_time_);
89 double first_chunk = profile_->integrate_simple(a - (a_index - 1) * last_time_, last_time_);
90 double middle_chunk = (b_index - a_index) * total_;
91 double last_chunk = profile_->integrate_simple(0.0, b - b_index * last_time_);
93 XBT_DEBUG("first_chunk=%.2f middle_chunk=%.2f last_chunk=%.2f\n", first_chunk, middle_chunk, last_chunk);
95 return (first_chunk + middle_chunk + last_chunk);
99 * @brief Auxiliary function to compute the integral between a and b.
100 * It simply computes the integrals at point a and b and returns the difference between them.
101 * @param a Initial point
102 * @param b Final point
104 double CpuTiProfile::integrate_simple(double a, double b) const
106 return integrate_simple_point(b) - integrate_simple_point(a);
110 * @brief Auxiliary function to compute the integral at point a.
113 double CpuTiProfile::integrate_simple_point(double a) const
117 long ind = binary_search(time_points_, a);
118 integral += integral_[ind];
120 XBT_DEBUG("a %f ind %ld integral %f ind + 1 %f ind %f time +1 %f time %f", a, ind, integral, integral_[ind + 1],
121 integral_[ind], time_points_[ind + 1], time_points_[ind]);
122 double_update(&a_aux, time_points_[ind], sg_precision_workamount * sg_precision_timing);
125 ((integral_[ind + 1] - integral_[ind]) / (time_points_[ind + 1] - time_points_[ind])) * (a - time_points_[ind]);
126 XBT_DEBUG("Integral a %f = %f", a, integral);
132 * @brief Computes the time needed to execute "amount" on cpu.
134 * Here, amount can span multiple trace periods
136 * @param a Initial time
137 * @param amount Amount to be executed
140 double CpuTiTmgr::solve(double a, double amount) const
142 /* Fix very small negative numbers */
143 if ((a < 0.0) && (a > -EPSILON)) {
146 if ((amount < 0.0) && (amount > -EPSILON)) {
151 xbt_assert(a >= 0.0 && amount >= 0.0,
152 "Error, invalid parameters [a = %.2f, amount = %.2f]. "
153 "You probably have a task executing with negative computation amount. Check your code.",
156 /* At this point, a and amount are positive */
157 if (amount < EPSILON)
160 /* Is the trace fixed ? */
161 if (type_ == Type::FIXED) {
162 return (a + (amount / value_));
165 XBT_DEBUG("amount %f total %f", amount, total_);
166 /* Reduce the problem to one where amount <= trace_total */
167 double quotient = floor(amount / total_);
168 double reduced_amount = total_ * ((amount / total_) - floor(amount / total_));
169 double reduced_a = a - last_time_ * static_cast<int>(floor(a / last_time_));
171 XBT_DEBUG("Quotient: %g reduced_amount: %f reduced_a: %f", quotient, reduced_amount, reduced_a);
173 /* Now solve for new_amount which is <= trace_total */
174 XBT_DEBUG("Solve integral: [%.2f, amount=%.2f]", reduced_a, reduced_amount);
176 double amount_till_end = integrate(reduced_a, last_time_);
177 double reduced_b = amount_till_end > reduced_amount
178 ? profile_->solve_simple(reduced_a, reduced_amount)
179 : last_time_ + profile_->solve_simple(0.0, reduced_amount - amount_till_end);
181 /* Re-map to the original b and amount */
182 return last_time_ * floor(a / last_time_) + (quotient * last_time_) + reduced_b;
186 * @brief Auxiliary function to solve integral.
187 * It returns the date when the requested amount of flops is available
188 * @param a Initial point
189 * @param amount Amount of flops
190 * @return The date when amount is available.
192 double CpuTiProfile::solve_simple(double a, double amount) const
194 double integral_a = integrate_simple_point(a);
195 long ind = binary_search(integral_, integral_a + amount);
196 double time = time_points_[ind];
197 time += (integral_a + amount - integral_[ind]) /
198 ((integral_[ind + 1] - integral_[ind]) / (time_points_[ind + 1] - time_points_[ind]));
204 * @brief Auxiliary function to update the CPU speed scale.
206 * This function uses the trace structure to return the speed scale at the determined time a.
208 * @return CPU speed scale
210 double CpuTiTmgr::get_power_scale(double a) const
212 double reduced_a = a - floor(a / last_time_) * last_time_;
213 long point = CpuTiProfile::binary_search(profile_->get_time_points(), reduced_a);
214 profile::DatedValue val = speed_profile_->get_event_list().at(point);
219 * @brief Creates a new integration trace from a tmgr_trace_t
221 * @param speed_trace CPU availability trace
222 * @param value Percentage of CPU speed available (useful to fixed tracing)
223 * @return Integration trace structure
225 CpuTiTmgr::CpuTiTmgr(kernel::profile::Profile* speed_profile, double value) : speed_profile_(speed_profile)
227 double total_time = 0.0;
228 profile_.reset(nullptr);
230 /* no availability file, fixed trace */
231 if (not speed_profile) {
233 XBT_DEBUG("No availability trace. Constant value = %f", value);
237 xbt_assert(speed_profile->is_repeating());
239 /* only one point available, fixed trace */
240 if (speed_profile->get_event_list().size() == 1) {
241 value_ = speed_profile->get_event_list().front().value_;
245 type_ = Type::DYNAMIC;
247 /* count the total time of trace file */
248 for (auto const& val : speed_profile->get_event_list())
249 total_time += val.date_;
250 total_time += speed_profile->get_repeat_delay();
252 profile_ = std::make_unique<CpuTiProfile>(speed_profile);
253 last_time_ = total_time;
254 total_ = profile_->integrate_simple(0, total_time);
256 XBT_DEBUG("Total integral %f, last_time %f ", total_, last_time_);
260 * @brief Binary search in array.
261 * It returns the last point of the interval in which "a" is.
263 * @param a Value to search
264 * @return Index of point
266 long CpuTiProfile::binary_search(const std::vector<double>& array, double a)
270 auto pos = std::upper_bound(begin(array), end(array), a);
271 return std::distance(begin(array), pos) - 1;
278 void CpuTiModel::create_pm_models()
280 auto cpu_model_pm = std::make_shared<CpuTiModel>("Cpu_TI");
281 auto* engine = EngineImpl::get_instance();
282 engine->add_model(cpu_model_pm);
283 engine->get_netzone_root()->set_cpu_pm_model(cpu_model_pm);
286 CpuImpl* CpuTiModel::create_cpu(s4u::Host* host, const std::vector<double>& speed_per_pstate)
288 return (new CpuTi(host, speed_per_pstate))->set_model(this);
291 double CpuTiModel::next_occurring_event(double now)
293 double min_action_duration = -1;
295 /* iterates over modified cpus to update share resources */
296 for (auto it = std::begin(modified_cpus_); it != std::end(modified_cpus_);) {
298 ++it; // increment iterator here since the following call to ti.update_actions_finish_time() may invalidate it
299 cpu.update_actions_finish_time(now);
302 /* get the min next event if heap not empty */
303 if (not get_action_heap().empty())
304 min_action_duration = get_action_heap().top_date() - now;
306 XBT_DEBUG("Share resources, min next event date: %f", min_action_duration);
308 return min_action_duration;
311 void CpuTiModel::update_actions_state(double now, double /*delta*/)
313 while (not get_action_heap().empty() && double_equals(get_action_heap().top_date(), now, sg_precision_timing)) {
314 auto* action = static_cast<CpuTiAction*>(get_action_heap().pop());
315 XBT_DEBUG("Action %p: finish", action);
316 action->finish(Action::State::FINISHED);
317 /* update remaining amount of all actions */
318 action->cpu_->update_remaining_amount(EngineImpl::get_clock());
325 CpuTi::CpuTi(s4u::Host* host, const std::vector<double>& speed_per_pstate) : CpuImpl(host, speed_per_pstate)
327 speed_.peak = speed_per_pstate.front();
328 XBT_DEBUG("CPU create: peak=%f", speed_.peak);
330 speed_integrated_trace_ = new CpuTiTmgr(nullptr, 1 /*scale*/);
336 delete speed_integrated_trace_;
339 CpuImpl* CpuTi::set_speed_profile(kernel::profile::Profile* profile)
341 delete speed_integrated_trace_;
342 speed_integrated_trace_ = new CpuTiTmgr(profile, speed_.scale);
344 /* add a fake trace event if periodicity == 0 */
345 if (profile && profile->get_event_list().size() > 1) {
346 kernel::profile::DatedValue val = profile->get_event_list().back();
347 if (val.date_ < 1e-12) {
348 auto* prof = profile::ProfileBuilder::from_void();
349 speed_.event = prof->schedule(&profile::future_evt_set, this);
355 void CpuTi::apply_event(kernel::profile::Event* event, double value)
357 if (event == speed_.event) {
358 XBT_DEBUG("Speed changed in trace! New fixed value: %f", value);
360 /* update remaining of actions and put in modified cpu list */
361 update_remaining_amount(EngineImpl::get_clock());
365 delete speed_integrated_trace_;
366 speed_integrated_trace_ = new CpuTiTmgr(value);
368 speed_.scale = value;
369 tmgr_trace_event_unref(&speed_.event);
371 } else if (event == get_state_event()) {
374 XBT_VERB("Restart actors on host %s", get_iface()->get_cname());
375 get_iface()->turn_on();
378 get_iface()->turn_off();
379 double date = EngineImpl::get_clock();
381 /* put all action running on cpu to failed */
382 for (CpuTiAction& action : action_set_) {
383 if (action.get_state() == Action::State::INITED || action.get_state() == Action::State::STARTED ||
384 action.get_state() == Action::State::IGNORED) {
385 action.set_finish_time(date);
386 action.set_state(Action::State::FAILED);
387 get_model()->get_action_heap().remove(&action);
394 xbt_die("Unknown event!\n");
398 /** Update the actions that are running on this CPU (which was modified recently) */
399 void CpuTi::update_actions_finish_time(double now)
401 /* update remaining amount of actions */
402 update_remaining_amount(now);
404 /* Compute the sum of priorities for the actions running on that CPU */
406 for (CpuTiAction const& action : action_set_) {
407 /* action not running, skip it */
408 if (action.get_state_set() != get_model()->get_started_action_set())
411 /* bogus priority, skip it */
412 if (action.get_sharing_penalty() <= 0)
415 /* action suspended, skip it */
416 if (not action.is_running())
419 sum_priority_ += 1.0 / action.get_sharing_penalty();
422 for (CpuTiAction& action : action_set_) {
423 double min_finish = NO_MAX_DURATION;
424 /* action not running, skip it */
425 if (action.get_state_set() != get_model()->get_started_action_set())
428 /* verify if the action is really running on cpu */
429 if (action.is_running() && action.get_sharing_penalty() > 0) {
430 /* total area needed to finish the action. Used in trace integration */
431 double total_area = (action.get_remains() * sum_priority_ * action.get_sharing_penalty()) / speed_.peak;
433 action.set_finish_time(speed_integrated_trace_->solve(now, total_area));
434 /* verify which event will happen before (max_duration or finish time) */
435 if (action.get_max_duration() != NO_MAX_DURATION &&
436 action.get_start_time() + action.get_max_duration() < action.get_finish_time())
437 min_finish = action.get_start_time() + action.get_max_duration();
439 min_finish = action.get_finish_time();
441 /* put the max duration time on heap */
442 if (action.get_max_duration() != NO_MAX_DURATION)
443 min_finish = action.get_start_time() + action.get_max_duration();
445 /* add in action heap */
446 if (min_finish != NO_MAX_DURATION)
447 get_model()->get_action_heap().update(&action, min_finish, ActionHeap::Type::unset);
449 get_model()->get_action_heap().remove(&action);
451 XBT_DEBUG("Update finish time: Cpu(%s) Action: %p, Start Time: %f Finish Time: %f Max duration %f", get_cname(),
452 &action, action.get_start_time(), action.get_finish_time(), action.get_max_duration());
454 /* remove from modified cpu */
458 bool CpuTi::is_used() const
460 return not action_set_.empty();
463 double CpuTi::get_speed_ratio()
465 speed_.scale = speed_integrated_trace_->get_power_scale(EngineImpl::get_clock());
466 return CpuImpl::get_speed_ratio();
469 /** @brief Update the remaining amount of actions */
470 void CpuTi::update_remaining_amount(double now)
472 /* already up to date */
473 if (last_update_ >= now)
476 /* compute the integration area */
477 double area_total = speed_integrated_trace_->integrate(last_update_, now) * speed_.peak;
478 XBT_DEBUG("Flops total: %f, Last update %f", area_total, last_update_);
479 for (CpuTiAction& action : action_set_) {
480 /* action not running, skip it */
481 if (action.get_state_set() != get_model()->get_started_action_set())
484 /* bogus priority, skip it */
485 if (action.get_sharing_penalty() <= 0)
488 /* action suspended, skip it */
489 if (not action.is_running())
492 /* action don't need update */
493 if (action.get_start_time() >= now)
496 /* skip action that are finishing now */
497 if (action.get_finish_time() >= 0 && action.get_finish_time() <= now)
500 /* update remaining */
501 action.update_remains(area_total / (sum_priority_ * action.get_sharing_penalty()));
502 XBT_DEBUG("Update remaining action(%p) remaining %f", &action, action.get_remains_no_update());
507 CpuAction* CpuTi::execution_start(double size, double user_bound)
509 XBT_IN("(%s,%g)", get_cname(), size);
510 xbt_assert(user_bound <= 0, "Invalid user bound (%lf) in CPU TI model", user_bound);
511 auto* action = new CpuTiAction(this, size);
513 action_set_.push_back(*action); // Actually start the action
519 CpuAction* CpuTi::sleep(double duration)
522 duration = std::max(duration, sg_precision_timing);
524 XBT_IN("(%s,%g)", get_cname(), duration);
525 auto* action = new CpuTiAction(this, 1.0);
527 action->set_max_duration(duration);
528 action->set_suspend_state(Action::SuspendStates::SLEEPING);
529 if (duration == NO_MAX_DURATION)
530 action->set_state(Action::State::IGNORED);
532 action_set_.push_back(*action);
538 void CpuTi::set_modified(bool modified)
540 CpuTiList& modified_cpus = static_cast<CpuTiModel*>(get_model())->modified_cpus_;
542 if (not cpu_ti_hook.is_linked()) {
543 modified_cpus.push_back(*this);
546 if (cpu_ti_hook.is_linked())
547 xbt::intrusive_erase(modified_cpus, *this);
555 CpuTiAction::CpuTiAction(CpuTi* cpu, double cost) : CpuAction(cpu->get_model(), cost, not cpu->is_on()), cpu_(cpu)
557 cpu_->set_modified(true);
559 CpuTiAction::~CpuTiAction()
561 /* remove from action_set */
562 if (action_ti_hook.is_linked())
563 xbt::intrusive_erase(cpu_->action_set_, *this);
564 /* remove from heap */
565 get_model()->get_action_heap().remove(this);
566 cpu_->set_modified(true);
569 void CpuTiAction::set_state(Action::State state)
571 CpuAction::set_state(state);
572 cpu_->set_modified(true);
575 void CpuTiAction::cancel()
577 this->set_state(Action::State::FAILED);
578 get_model()->get_action_heap().remove(this);
579 cpu_->set_modified(true);
582 void CpuTiAction::suspend()
584 XBT_IN("(%p)", this);
586 set_suspend_state(Action::SuspendStates::SUSPENDED);
587 get_model()->get_action_heap().remove(this);
588 cpu_->set_modified(true);
593 void CpuTiAction::resume()
595 XBT_IN("(%p)", this);
596 if (is_suspended()) {
597 set_suspend_state(Action::SuspendStates::RUNNING);
598 cpu_->set_modified(true);
603 void CpuTiAction::set_sharing_penalty(double sharing_penalty)
605 XBT_IN("(%p,%g)", this, sharing_penalty);
606 set_sharing_penalty_no_update(sharing_penalty);
607 cpu_->set_modified(true);
611 double CpuTiAction::get_remains()
613 XBT_IN("(%p)", this);
614 cpu_->update_remaining_amount(EngineImpl::get_clock());
616 return get_remains_no_update();
619 } // namespace simgrid::kernel::resource