1 /* Copyright (c) 2013-2022. 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 "xbt/asserts.h"
10 #include "src/kernel/EngineImpl.hpp"
11 #include "src/kernel/resource/profile/Event.hpp"
12 #include "src/kernel/resource/profile/Profile.hpp"
13 #include "src/surf/surf_interface.hpp"
18 constexpr double EPSILON = 0.000000001;
20 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(cpu_ti, res_cpu, "CPU resource, Trace Integration model");
30 CpuTiProfile::CpuTiProfile(const profile::Profile* profile)
34 double prev_value = 1;
35 const std::vector<profile::DatedValue>& events=profile->get_event_list();
36 xbt_assert(not events.empty());
37 unsigned long nb_points = events.size() + 1;
38 time_points_.reserve(nb_points);
39 integral_.reserve(nb_points);
40 for (auto const& val : events) {
42 integral += val.date_ * prev_value;
43 time_points_.push_back(time);
44 integral_.push_back(integral);
45 prev_value = val.value_;
48 double delay=profile->get_repeat_delay()+ events.at(0).date_;
50 xbt_assert( events.back().value_==prev_value,"Profiles need to end as they start");
52 integral += delay*prev_value;
54 time_points_.push_back(time);
55 integral_.push_back(integral);
59 * @brief Integrate trace
61 * Wrapper around surf_cpu_integrate_trace_simple() to get
64 * @param a Begin of interval
65 * @param b End of interval
66 * @return the integrate value. -1 if an error occurs.
68 double CpuTiTmgr::integrate(double a, double b) const
70 xbt_assert(a >= 0.0 && a <= b,
71 "Error, invalid integration interval [%.2f,%.2f]. You probably have a task executing with negative "
72 "computation amount. Check your code.",
74 if (fabs(a - b) < EPSILON)
77 if (type_ == Type::FIXED) {
78 return (b - a) * value_;
82 if (fabs(ceil(a / last_time_) - a / last_time_) < EPSILON)
83 a_index = 1 + ceil(a / last_time_);
85 a_index = ceil(a / last_time_);
86 double b_index = floor(b / last_time_);
88 if (a_index > b_index) { /* Same chunk */
89 return profile_->integrate_simple(a - (a_index - 1) * last_time_, b - b_index * last_time_);
92 double first_chunk = profile_->integrate_simple(a - (a_index - 1) * last_time_, last_time_);
93 double middle_chunk = (b_index - a_index) * total_;
94 double last_chunk = profile_->integrate_simple(0.0, b - b_index * last_time_);
96 XBT_DEBUG("first_chunk=%.2f middle_chunk=%.2f last_chunk=%.2f\n", first_chunk, middle_chunk, last_chunk);
98 return (first_chunk + middle_chunk + last_chunk);
102 * @brief Auxiliary function to compute the integral between a and b.
103 * It simply computes the integrals at point a and b and returns the difference between them.
104 * @param a Initial point
105 * @param b Final point
107 double CpuTiProfile::integrate_simple(double a, double b) const
109 return integrate_simple_point(b) - integrate_simple_point(a);
113 * @brief Auxiliary function to compute the integral at point a.
116 double CpuTiProfile::integrate_simple_point(double a) const
120 long ind = binary_search(time_points_, a);
121 integral += integral_[ind];
123 XBT_DEBUG("a %f ind %ld integral %f ind + 1 %f ind %f time +1 %f time %f", a, ind, integral, integral_[ind + 1],
124 integral_[ind], time_points_[ind + 1], time_points_[ind]);
125 double_update(&a_aux, time_points_[ind], sg_maxmin_precision * sg_surf_precision);
128 ((integral_[ind + 1] - integral_[ind]) / (time_points_[ind + 1] - time_points_[ind])) * (a - time_points_[ind]);
129 XBT_DEBUG("Integral a %f = %f", a, integral);
135 * @brief Computes the time needed to execute "amount" on cpu.
137 * Here, amount can span multiple trace periods
139 * @param a Initial time
140 * @param amount Amount to be executed
143 double CpuTiTmgr::solve(double a, double amount) const
145 /* Fix very small negative numbers */
146 if ((a < 0.0) && (a > -EPSILON)) {
149 if ((amount < 0.0) && (amount > -EPSILON)) {
154 xbt_assert(a >= 0.0 && amount >= 0.0,
155 "Error, invalid parameters [a = %.2f, amount = %.2f]. "
156 "You probably have a task executing with negative computation amount. Check your code.",
159 /* At this point, a and amount are positive */
160 if (amount < EPSILON)
163 /* Is the trace fixed ? */
164 if (type_ == Type::FIXED) {
165 return (a + (amount / value_));
168 XBT_DEBUG("amount %f total %f", amount, total_);
169 /* Reduce the problem to one where amount <= trace_total */
170 double quotient = floor(amount / total_);
171 double reduced_amount = total_ * ((amount / total_) - floor(amount / total_));
172 double reduced_a = a - last_time_ * static_cast<int>(floor(a / last_time_));
174 XBT_DEBUG("Quotient: %g reduced_amount: %f reduced_a: %f", quotient, reduced_amount, reduced_a);
176 /* Now solve for new_amount which is <= trace_total */
177 XBT_DEBUG("Solve integral: [%.2f, amount=%.2f]", reduced_a, reduced_amount);
179 double amount_till_end = integrate(reduced_a, last_time_);
180 double reduced_b = amount_till_end > reduced_amount
181 ? profile_->solve_simple(reduced_a, reduced_amount)
182 : last_time_ + profile_->solve_simple(0.0, reduced_amount - amount_till_end);
184 /* Re-map to the original b and amount */
185 return last_time_ * floor(a / last_time_) + (quotient * last_time_) + reduced_b;
189 * @brief Auxiliary function to solve integral.
190 * It returns the date when the requested amount of flops is available
191 * @param a Initial point
192 * @param amount Amount of flops
193 * @return The date when amount is available.
195 double CpuTiProfile::solve_simple(double a, double amount) const
197 double integral_a = integrate_simple_point(a);
198 long ind = binary_search(integral_, integral_a + amount);
199 double time = time_points_[ind];
200 time += (integral_a + amount - integral_[ind]) /
201 ((integral_[ind + 1] - integral_[ind]) / (time_points_[ind + 1] - time_points_[ind]));
207 * @brief Auxiliary function to update the CPU speed scale.
209 * This function uses the trace structure to return the speed scale at the determined time a.
211 * @return CPU speed scale
213 double CpuTiTmgr::get_power_scale(double a) const
215 double reduced_a = a - floor(a / last_time_) * last_time_;
216 long point = CpuTiProfile::binary_search(profile_->get_time_points(), reduced_a);
217 profile::DatedValue val = speed_profile_->get_event_list().at(point);
222 * @brief Creates a new integration trace from a tmgr_trace_t
224 * @param speed_trace CPU availability trace
225 * @param value Percentage of CPU speed available (useful to fixed tracing)
226 * @return Integration trace structure
228 CpuTiTmgr::CpuTiTmgr(kernel::profile::Profile* speed_profile, double value) : speed_profile_(speed_profile)
230 double total_time = 0.0;
231 profile_.reset(nullptr);
233 /* no availability file, fixed trace */
234 if (not speed_profile) {
236 XBT_DEBUG("No availability trace. Constant value = %f", value);
240 xbt_assert(speed_profile->is_repeating());
242 /* only one point available, fixed trace */
243 if (speed_profile->get_event_list().size() == 1) {
244 value_ = speed_profile->get_event_list().front().value_;
248 type_ = Type::DYNAMIC;
250 /* count the total time of trace file */
251 for (auto const& val : speed_profile->get_event_list())
252 total_time += val.date_;
253 total_time += speed_profile->get_repeat_delay();
255 profile_ = std::make_unique<CpuTiProfile>(speed_profile);
256 last_time_ = total_time;
257 total_ = profile_->integrate_simple(0, total_time);
259 XBT_DEBUG("Total integral %f, last_time %f ", total_, last_time_);
263 * @brief Binary search in array.
264 * It returns the last point of the interval in which "a" is.
266 * @param a Value to search
267 * @return Index of point
269 long CpuTiProfile::binary_search(const std::vector<double>& array, double a)
273 auto pos = std::upper_bound(begin(array), end(array), a);
274 return std::distance(begin(array), pos) - 1;
281 void CpuTiModel::create_pm_models()
283 auto cpu_model_pm = std::make_shared<CpuTiModel>("Cpu_TI");
284 auto* engine = EngineImpl::get_instance();
285 engine->add_model(cpu_model_pm);
286 engine->get_netzone_root()->set_cpu_pm_model(cpu_model_pm);
289 CpuImpl* CpuTiModel::create_cpu(s4u::Host* host, const std::vector<double>& speed_per_pstate)
291 return (new CpuTi(host, speed_per_pstate))->set_model(this);
294 double CpuTiModel::next_occurring_event(double now)
296 double min_action_duration = -1;
298 /* iterates over modified cpus to update share resources */
299 for (auto it = std::begin(modified_cpus_); it != std::end(modified_cpus_);) {
301 ++it; // increment iterator here since the following call to ti.update_actions_finish_time() may invalidate it
302 cpu.update_actions_finish_time(now);
305 /* get the min next event if heap not empty */
306 if (not get_action_heap().empty())
307 min_action_duration = get_action_heap().top_date() - now;
309 XBT_DEBUG("Share resources, min next event date: %f", min_action_duration);
311 return min_action_duration;
314 void CpuTiModel::update_actions_state(double now, double /*delta*/)
316 while (not get_action_heap().empty() && double_equals(get_action_heap().top_date(), now, sg_surf_precision)) {
317 auto* action = static_cast<CpuTiAction*>(get_action_heap().pop());
318 XBT_DEBUG("Action %p: finish", action);
319 action->finish(Action::State::FINISHED);
320 /* update remaining amount of all actions */
321 action->cpu_->update_remaining_amount(EngineImpl::get_clock());
328 CpuTi::CpuTi(s4u::Host* host, const std::vector<double>& speed_per_pstate) : CpuImpl(host, speed_per_pstate)
330 speed_.peak = speed_per_pstate.front();
331 XBT_DEBUG("CPU create: peak=%f", speed_.peak);
333 speed_integrated_trace_ = new CpuTiTmgr(nullptr, 1 /*scale*/);
339 delete speed_integrated_trace_;
342 CpuImpl* CpuTi::set_speed_profile(kernel::profile::Profile* profile)
344 delete speed_integrated_trace_;
345 speed_integrated_trace_ = new CpuTiTmgr(profile, speed_.scale);
347 /* add a fake trace event if periodicity == 0 */
348 if (profile && profile->get_event_list().size() > 1) {
349 kernel::profile::DatedValue val = profile->get_event_list().back();
350 if (val.date_ < 1e-12) {
351 auto* prof = profile::ProfileBuilder::from_void();
352 speed_.event = prof->schedule(&profile::future_evt_set, this);
358 void CpuTi::apply_event(kernel::profile::Event* event, double value)
360 if (event == speed_.event) {
361 XBT_DEBUG("Speed changed in trace! New fixed value: %f", value);
363 /* update remaining of actions and put in modified cpu list */
364 update_remaining_amount(EngineImpl::get_clock());
368 delete speed_integrated_trace_;
369 speed_integrated_trace_ = new CpuTiTmgr(value);
371 speed_.scale = value;
372 tmgr_trace_event_unref(&speed_.event);
374 } else if (event == get_state_event()) {
377 XBT_VERB("Restart actors on host %s", get_iface()->get_cname());
378 get_iface()->turn_on();
381 get_iface()->turn_off();
382 double date = EngineImpl::get_clock();
384 /* put all action running on cpu to failed */
385 for (CpuTiAction& action : action_set_) {
386 if (action.get_state() == Action::State::INITED || action.get_state() == Action::State::STARTED ||
387 action.get_state() == Action::State::IGNORED) {
388 action.set_finish_time(date);
389 action.set_state(Action::State::FAILED);
390 get_model()->get_action_heap().remove(&action);
397 xbt_die("Unknown event!\n");
401 /** Update the actions that are running on this CPU (which was modified recently) */
402 void CpuTi::update_actions_finish_time(double now)
404 /* update remaining amount of actions */
405 update_remaining_amount(now);
407 /* Compute the sum of priorities for the actions running on that CPU */
409 for (CpuTiAction const& action : action_set_) {
410 /* action not running, skip it */
411 if (action.get_state_set() != get_model()->get_started_action_set())
414 /* bogus priority, skip it */
415 if (action.get_sharing_penalty() <= 0)
418 /* action suspended, skip it */
419 if (not action.is_running())
422 sum_priority_ += 1.0 / action.get_sharing_penalty();
425 for (CpuTiAction& action : action_set_) {
426 double min_finish = NO_MAX_DURATION;
427 /* action not running, skip it */
428 if (action.get_state_set() != get_model()->get_started_action_set())
431 /* verify if the action is really running on cpu */
432 if (action.is_running() && action.get_sharing_penalty() > 0) {
433 /* total area needed to finish the action. Used in trace integration */
434 double total_area = (action.get_remains() * sum_priority_ * action.get_sharing_penalty()) / speed_.peak;
436 action.set_finish_time(speed_integrated_trace_->solve(now, total_area));
437 /* verify which event will happen before (max_duration or finish time) */
438 if (action.get_max_duration() != NO_MAX_DURATION &&
439 action.get_start_time() + action.get_max_duration() < action.get_finish_time())
440 min_finish = action.get_start_time() + action.get_max_duration();
442 min_finish = action.get_finish_time();
444 /* put the max duration time on heap */
445 if (action.get_max_duration() != NO_MAX_DURATION)
446 min_finish = action.get_start_time() + action.get_max_duration();
448 /* add in action heap */
449 if (min_finish != NO_MAX_DURATION)
450 get_model()->get_action_heap().update(&action, min_finish, ActionHeap::Type::unset);
452 get_model()->get_action_heap().remove(&action);
454 XBT_DEBUG("Update finish time: Cpu(%s) Action: %p, Start Time: %f Finish Time: %f Max duration %f", get_cname(),
455 &action, action.get_start_time(), action.get_finish_time(), action.get_max_duration());
457 /* remove from modified cpu */
461 bool CpuTi::is_used() const
463 return not action_set_.empty();
466 double CpuTi::get_speed_ratio()
468 speed_.scale = speed_integrated_trace_->get_power_scale(EngineImpl::get_clock());
469 return CpuImpl::get_speed_ratio();
472 /** @brief Update the remaining amount of actions */
473 void CpuTi::update_remaining_amount(double now)
475 /* already up to date */
476 if (last_update_ >= now)
479 /* compute the integration area */
480 double area_total = speed_integrated_trace_->integrate(last_update_, now) * speed_.peak;
481 XBT_DEBUG("Flops total: %f, Last update %f", area_total, last_update_);
482 for (CpuTiAction& action : action_set_) {
483 /* action not running, skip it */
484 if (action.get_state_set() != get_model()->get_started_action_set())
487 /* bogus priority, skip it */
488 if (action.get_sharing_penalty() <= 0)
491 /* action suspended, skip it */
492 if (not action.is_running())
495 /* action don't need update */
496 if (action.get_start_time() >= now)
499 /* skip action that are finishing now */
500 if (action.get_finish_time() >= 0 && action.get_finish_time() <= now)
503 /* update remaining */
504 action.update_remains(area_total / (sum_priority_ * action.get_sharing_penalty()));
505 XBT_DEBUG("Update remaining action(%p) remaining %f", &action, action.get_remains_no_update());
510 CpuAction* CpuTi::execution_start(double size, double user_bound)
512 XBT_IN("(%s,%g)", get_cname(), size);
513 xbt_assert(user_bound <= 0, "Invalid user bound (%lf) in CPU TI model", user_bound);
514 auto* action = new CpuTiAction(this, size);
516 action_set_.push_back(*action); // Actually start the action
522 CpuAction* CpuTi::sleep(double duration)
525 duration = std::max(duration, sg_surf_precision);
527 XBT_IN("(%s,%g)", get_cname(), duration);
528 auto* action = new CpuTiAction(this, 1.0);
530 action->set_max_duration(duration);
531 action->set_suspend_state(Action::SuspendStates::SLEEPING);
532 if (duration == NO_MAX_DURATION)
533 action->set_state(Action::State::IGNORED);
535 action_set_.push_back(*action);
541 void CpuTi::set_modified(bool modified)
543 CpuTiList& modified_cpus = static_cast<CpuTiModel*>(get_model())->modified_cpus_;
545 if (not cpu_ti_hook.is_linked()) {
546 modified_cpus.push_back(*this);
549 if (cpu_ti_hook.is_linked())
550 xbt::intrusive_erase(modified_cpus, *this);
558 CpuTiAction::CpuTiAction(CpuTi* cpu, double cost) : CpuAction(cpu->get_model(), cost, not cpu->is_on()), cpu_(cpu)
560 cpu_->set_modified(true);
562 CpuTiAction::~CpuTiAction()
564 /* remove from action_set */
565 if (action_ti_hook.is_linked())
566 xbt::intrusive_erase(cpu_->action_set_, *this);
567 /* remove from heap */
568 get_model()->get_action_heap().remove(this);
569 cpu_->set_modified(true);
572 void CpuTiAction::set_state(Action::State state)
574 CpuAction::set_state(state);
575 cpu_->set_modified(true);
578 void CpuTiAction::cancel()
580 this->set_state(Action::State::FAILED);
581 get_model()->get_action_heap().remove(this);
582 cpu_->set_modified(true);
585 void CpuTiAction::suspend()
587 XBT_IN("(%p)", this);
589 set_suspend_state(Action::SuspendStates::SUSPENDED);
590 get_model()->get_action_heap().remove(this);
591 cpu_->set_modified(true);
596 void CpuTiAction::resume()
598 XBT_IN("(%p)", this);
599 if (is_suspended()) {
600 set_suspend_state(Action::SuspendStates::RUNNING);
601 cpu_->set_modified(true);
606 void CpuTiAction::set_sharing_penalty(double sharing_penalty)
608 XBT_IN("(%p,%g)", this, sharing_penalty);
609 set_sharing_penalty_no_update(sharing_penalty);
610 cpu_->set_modified(true);
614 double CpuTiAction::get_remains()
616 XBT_IN("(%p)", this);
617 cpu_->update_remaining_amount(EngineImpl::get_clock());
619 return get_remains_no_update();
622 } // namespace resource
623 } // namespace kernel
624 } // namespace simgrid