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 "src/kernel/EngineImpl.hpp"
10 #include "src/kernel/resource/profile/Event.hpp"
11 #include "src/kernel/resource/profile/Profile.hpp"
12 #include "src/surf/surf_interface.hpp"
17 constexpr double EPSILON = 0.000000001;
19 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(cpu_ti, res_cpu, "CPU resource, Trace Integration model");
29 CpuTiProfile::CpuTiProfile(const profile::Profile* profile)
33 unsigned long nb_points = profile->get_event_list().size() + 1;
34 time_points_.reserve(nb_points);
35 integral_.reserve(nb_points);
36 for (auto const& val : profile->get_event_list()) {
37 time_points_.push_back(time);
38 integral_.push_back(integral);
40 integral += val.date_ * val.value_;
42 time_points_.push_back(time);
43 integral_.push_back(integral);
47 * @brief Integrate trace
49 * Wrapper around surf_cpu_integrate_trace_simple() to get
52 * @param a Begin of interval
53 * @param b End of interval
54 * @return the integrate value. -1 if an error occurs.
56 double CpuTiTmgr::integrate(double a, double b) const
58 xbt_assert(a >= 0.0 && a <= b,
59 "Error, invalid integration interval [%.2f,%.2f]. You probably have a task executing with negative "
60 "computation amount. Check your code.",
62 if (fabs(a - b) < EPSILON)
65 if (type_ == Type::FIXED) {
66 return (b - a) * value_;
70 if (fabs(ceil(a / last_time_) - a / last_time_) < EPSILON)
71 a_index = 1 + ceil(a / last_time_);
73 a_index = ceil(a / last_time_);
74 double b_index = floor(b / last_time_);
76 if (a_index > b_index) { /* Same chunk */
77 return profile_->integrate_simple(a - (a_index - 1) * last_time_, b - b_index * last_time_);
80 double first_chunk = profile_->integrate_simple(a - (a_index - 1) * last_time_, last_time_);
81 double middle_chunk = (b_index - a_index) * total_;
82 double last_chunk = profile_->integrate_simple(0.0, b - b_index * last_time_);
84 XBT_DEBUG("first_chunk=%.2f middle_chunk=%.2f last_chunk=%.2f\n", first_chunk, middle_chunk, last_chunk);
86 return (first_chunk + middle_chunk + last_chunk);
90 * @brief Auxiliary function to compute the integral between a and b.
91 * It simply computes the integrals at point a and b and returns the difference between them.
92 * @param a Initial point
93 * @param b Final point
95 double CpuTiProfile::integrate_simple(double a, double b) const
97 return integrate_simple_point(b) - integrate_simple_point(a);
101 * @brief Auxiliary function to compute the integral at point a.
104 double CpuTiProfile::integrate_simple_point(double a) const
108 long ind = binary_search(time_points_, a);
109 integral += integral_[ind];
111 XBT_DEBUG("a %f ind %ld integral %f ind + 1 %f ind %f time +1 %f time %f", a, ind, integral, integral_[ind + 1],
112 integral_[ind], time_points_[ind + 1], time_points_[ind]);
113 double_update(&a_aux, time_points_[ind], sg_maxmin_precision * sg_surf_precision);
116 ((integral_[ind + 1] - integral_[ind]) / (time_points_[ind + 1] - time_points_[ind])) * (a - time_points_[ind]);
117 XBT_DEBUG("Integral a %f = %f", a, integral);
123 * @brief Computes the time needed to execute "amount" on cpu.
125 * Here, amount can span multiple trace periods
127 * @param a Initial time
128 * @param amount Amount to be executed
131 double CpuTiTmgr::solve(double a, double amount) const
133 /* Fix very small negative numbers */
134 if ((a < 0.0) && (a > -EPSILON)) {
137 if ((amount < 0.0) && (amount > -EPSILON)) {
142 xbt_assert(a >= 0.0 && amount >= 0.0,
143 "Error, invalid parameters [a = %.2f, amount = %.2f]. "
144 "You probably have a task executing with negative computation amount. Check your code.",
147 /* At this point, a and amount are positive */
148 if (amount < EPSILON)
151 /* Is the trace fixed ? */
152 if (type_ == Type::FIXED) {
153 return (a + (amount / value_));
156 XBT_DEBUG("amount %f total %f", amount, total_);
157 /* Reduce the problem to one where amount <= trace_total */
158 double quotient = floor(amount / total_);
159 double reduced_amount = total_ * ((amount / total_) - floor(amount / total_));
160 double reduced_a = a - last_time_ * static_cast<int>(floor(a / last_time_));
162 XBT_DEBUG("Quotient: %g reduced_amount: %f reduced_a: %f", quotient, reduced_amount, reduced_a);
164 /* Now solve for new_amount which is <= trace_total */
166 XBT_DEBUG("Solve integral: [%.2f, amount=%.2f]", reduced_a, reduced_amount);
167 double amount_till_end = integrate(reduced_a, last_time_);
169 if (amount_till_end > reduced_amount) {
170 reduced_b = profile_->solve_simple(reduced_a, reduced_amount);
172 reduced_b = last_time_ + profile_->solve_simple(0.0, reduced_amount - amount_till_end);
175 /* Re-map to the original b and amount */
176 return last_time_ * floor(a / last_time_) + (quotient * last_time_) + reduced_b;
180 * @brief Auxiliary function to solve integral.
181 * It returns the date when the requested amount of flops is available
182 * @param a Initial point
183 * @param amount Amount of flops
184 * @return The date when amount is available.
186 double CpuTiProfile::solve_simple(double a, double amount) const
188 double integral_a = integrate_simple_point(a);
189 long ind = binary_search(integral_, integral_a + amount);
190 double time = time_points_[ind];
191 time += (integral_a + amount - integral_[ind]) /
192 ((integral_[ind + 1] - integral_[ind]) / (time_points_[ind + 1] - time_points_[ind]));
198 * @brief Auxiliary function to update the CPU speed scale.
200 * This function uses the trace structure to return the speed scale at the determined time a.
202 * @return CPU speed scale
204 double CpuTiTmgr::get_power_scale(double a) const
206 double reduced_a = a - floor(a / last_time_) * last_time_;
207 long point = CpuTiProfile::binary_search(profile_->get_time_points(), reduced_a);
208 profile::DatedValue val = speed_profile_->get_event_list().at(point);
213 * @brief Creates a new integration trace from a tmgr_trace_t
215 * @param speed_trace CPU availability trace
216 * @param value Percentage of CPU speed available (useful to fixed tracing)
217 * @return Integration trace structure
219 CpuTiTmgr::CpuTiTmgr(kernel::profile::Profile* speed_profile, double value) : speed_profile_(speed_profile)
221 double total_time = 0.0;
222 profile_.reset(nullptr);
224 /* no availability file, fixed trace */
225 if (not speed_profile) {
227 XBT_DEBUG("No availability trace. Constant value = %f", value);
231 /* only one point available, fixed trace */
232 if (speed_profile->get_event_list().size() == 1) {
233 value_ = speed_profile->get_event_list().front().value_;
237 type_ = Type::DYNAMIC;
239 /* count the total time of trace file */
240 for (auto const& val : speed_profile->get_event_list())
241 total_time += val.date_;
243 profile_ = std::make_unique<CpuTiProfile>(speed_profile);
244 last_time_ = total_time;
245 total_ = profile_->integrate_simple(0, total_time);
247 XBT_DEBUG("Total integral %f, last_time %f ", total_, last_time_);
251 * @brief Binary search in array.
252 * It returns the last point of the interval in which "a" is.
254 * @param a Value to search
255 * @return Index of point
257 long CpuTiProfile::binary_search(const std::vector<double>& array, double a)
261 auto pos = std::upper_bound(begin(array), end(array), a);
262 return std::distance(begin(array), pos) - 1;
269 void CpuTiModel::create_pm_models()
271 auto cpu_model_pm = std::make_shared<CpuTiModel>("Cpu_TI");
272 auto* engine = EngineImpl::get_instance();
273 engine->add_model(cpu_model_pm);
274 engine->get_netzone_root()->set_cpu_pm_model(cpu_model_pm);
277 CpuImpl* CpuTiModel::create_cpu(s4u::Host* host, const std::vector<double>& speed_per_pstate)
279 return (new CpuTi(host, speed_per_pstate))->set_model(this);
282 double CpuTiModel::next_occurring_event(double now)
284 double min_action_duration = -1;
286 /* iterates over modified cpus to update share resources */
287 for (auto it = std::begin(modified_cpus_); it != std::end(modified_cpus_);) {
289 ++it; // increment iterator here since the following call to ti.update_actions_finish_time() may invalidate it
290 cpu.update_actions_finish_time(now);
293 /* get the min next event if heap not empty */
294 if (not get_action_heap().empty())
295 min_action_duration = get_action_heap().top_date() - now;
297 XBT_DEBUG("Share resources, min next event date: %f", min_action_duration);
299 return min_action_duration;
302 void CpuTiModel::update_actions_state(double now, double /*delta*/)
304 while (not get_action_heap().empty() && double_equals(get_action_heap().top_date(), now, sg_surf_precision)) {
305 auto* action = static_cast<CpuTiAction*>(get_action_heap().pop());
306 XBT_DEBUG("Action %p: finish", action);
307 action->finish(Action::State::FINISHED);
308 /* update remaining amount of all actions */
309 action->cpu_->update_remaining_amount(EngineImpl::get_clock());
316 CpuTi::CpuTi(s4u::Host* host, const std::vector<double>& speed_per_pstate) : CpuImpl(host, speed_per_pstate)
318 speed_.peak = speed_per_pstate.front();
319 XBT_DEBUG("CPU create: peak=%f", speed_.peak);
321 speed_integrated_trace_ = new CpuTiTmgr(nullptr, 1 /*scale*/);
327 delete speed_integrated_trace_;
330 CpuImpl* CpuTi::set_speed_profile(kernel::profile::Profile* profile)
332 delete speed_integrated_trace_;
333 speed_integrated_trace_ = new CpuTiTmgr(profile, speed_.scale);
335 /* add a fake trace event if periodicity == 0 */
336 if (profile && profile->get_event_list().size() > 1) {
337 kernel::profile::DatedValue val = profile->get_event_list().back();
338 if (val.date_ < 1e-12) {
339 auto* prof = new kernel::profile::Profile();
340 speed_.event = prof->schedule(&profile::future_evt_set, this);
346 void CpuTi::apply_event(kernel::profile::Event* event, double value)
348 if (event == speed_.event) {
349 XBT_DEBUG("Speed changed in trace! New fixed value: %f", value);
351 /* update remaining of actions and put in modified cpu list */
352 update_remaining_amount(EngineImpl::get_clock());
356 delete speed_integrated_trace_;
357 speed_integrated_trace_ = new CpuTiTmgr(value);
359 speed_.scale = value;
360 tmgr_trace_event_unref(&speed_.event);
362 } else if (event == get_state_event()) {
365 XBT_VERB("Restart actors on host %s", get_iface()->get_cname());
366 get_iface()->turn_on();
369 get_iface()->turn_off();
370 double date = EngineImpl::get_clock();
372 /* put all action running on cpu to failed */
373 for (CpuTiAction& action : action_set_) {
374 if (action.get_state() == Action::State::INITED || action.get_state() == Action::State::STARTED ||
375 action.get_state() == Action::State::IGNORED) {
376 action.set_finish_time(date);
377 action.set_state(Action::State::FAILED);
378 get_model()->get_action_heap().remove(&action);
385 xbt_die("Unknown event!\n");
389 /** Update the actions that are running on this CPU (which was modified recently) */
390 void CpuTi::update_actions_finish_time(double now)
392 /* update remaining amount of actions */
393 update_remaining_amount(now);
395 /* Compute the sum of priorities for the actions running on that CPU */
397 for (CpuTiAction const& action : action_set_) {
398 /* action not running, skip it */
399 if (action.get_state_set() != get_model()->get_started_action_set())
402 /* bogus priority, skip it */
403 if (action.get_sharing_penalty() <= 0)
406 /* action suspended, skip it */
407 if (not action.is_running())
410 sum_priority_ += 1.0 / action.get_sharing_penalty();
413 for (CpuTiAction& action : action_set_) {
414 double min_finish = NO_MAX_DURATION;
415 /* action not running, skip it */
416 if (action.get_state_set() != get_model()->get_started_action_set())
419 /* verify if the action is really running on cpu */
420 if (action.is_running() && action.get_sharing_penalty() > 0) {
421 /* total area needed to finish the action. Used in trace integration */
422 double total_area = (action.get_remains() * sum_priority_ * action.get_sharing_penalty()) / speed_.peak;
424 action.set_finish_time(speed_integrated_trace_->solve(now, total_area));
425 /* verify which event will happen before (max_duration or finish time) */
426 if (action.get_max_duration() != NO_MAX_DURATION &&
427 action.get_start_time() + action.get_max_duration() < action.get_finish_time())
428 min_finish = action.get_start_time() + action.get_max_duration();
430 min_finish = action.get_finish_time();
432 /* put the max duration time on heap */
433 if (action.get_max_duration() != NO_MAX_DURATION)
434 min_finish = action.get_start_time() + action.get_max_duration();
436 /* add in action heap */
437 if (min_finish != NO_MAX_DURATION)
438 get_model()->get_action_heap().update(&action, min_finish, ActionHeap::Type::unset);
440 get_model()->get_action_heap().remove(&action);
442 XBT_DEBUG("Update finish time: Cpu(%s) Action: %p, Start Time: %f Finish Time: %f Max duration %f", get_cname(),
443 &action, action.get_start_time(), action.get_finish_time(), action.get_max_duration());
445 /* remove from modified cpu */
449 bool CpuTi::is_used() const
451 return not action_set_.empty();
454 double CpuTi::get_speed_ratio()
456 speed_.scale = speed_integrated_trace_->get_power_scale(EngineImpl::get_clock());
457 return CpuImpl::get_speed_ratio();
460 /** @brief Update the remaining amount of actions */
461 void CpuTi::update_remaining_amount(double now)
463 /* already up to date */
464 if (last_update_ >= now)
467 /* compute the integration area */
468 double area_total = speed_integrated_trace_->integrate(last_update_, now) * speed_.peak;
469 XBT_DEBUG("Flops total: %f, Last update %f", area_total, last_update_);
470 for (CpuTiAction& action : action_set_) {
471 /* action not running, skip it */
472 if (action.get_state_set() != get_model()->get_started_action_set())
475 /* bogus priority, skip it */
476 if (action.get_sharing_penalty() <= 0)
479 /* action suspended, skip it */
480 if (not action.is_running())
483 /* action don't need update */
484 if (action.get_start_time() >= now)
487 /* skip action that are finishing now */
488 if (action.get_finish_time() >= 0 && action.get_finish_time() <= now)
491 /* update remaining */
492 action.update_remains(area_total / (sum_priority_ * action.get_sharing_penalty()));
493 XBT_DEBUG("Update remaining action(%p) remaining %f", &action, action.get_remains_no_update());
498 CpuAction* CpuTi::execution_start(double size, double user_bound)
500 XBT_IN("(%s,%g)", get_cname(), size);
501 xbt_assert(user_bound <= 0, "Invalid user bound (%lf) in CPU TI model", user_bound);
502 auto* action = new CpuTiAction(this, size);
504 action_set_.push_back(*action); // Actually start the action
510 CpuAction* CpuTi::sleep(double duration)
513 duration = std::max(duration, sg_surf_precision);
515 XBT_IN("(%s,%g)", get_cname(), duration);
516 auto* action = new CpuTiAction(this, 1.0);
518 action->set_max_duration(duration);
519 action->set_suspend_state(Action::SuspendStates::SLEEPING);
520 if (duration == NO_MAX_DURATION)
521 action->set_state(Action::State::IGNORED);
523 action_set_.push_back(*action);
529 void CpuTi::set_modified(bool modified)
531 CpuTiList& modified_cpus = static_cast<CpuTiModel*>(get_model())->modified_cpus_;
533 if (not cpu_ti_hook.is_linked()) {
534 modified_cpus.push_back(*this);
537 if (cpu_ti_hook.is_linked())
538 xbt::intrusive_erase(modified_cpus, *this);
546 CpuTiAction::CpuTiAction(CpuTi* cpu, double cost) : CpuAction(cpu->get_model(), cost, not cpu->is_on()), cpu_(cpu)
548 cpu_->set_modified(true);
550 CpuTiAction::~CpuTiAction()
552 /* remove from action_set */
553 if (action_ti_hook.is_linked())
554 xbt::intrusive_erase(cpu_->action_set_, *this);
555 /* remove from heap */
556 get_model()->get_action_heap().remove(this);
557 cpu_->set_modified(true);
560 void CpuTiAction::set_state(Action::State state)
562 CpuAction::set_state(state);
563 cpu_->set_modified(true);
566 void CpuTiAction::cancel()
568 this->set_state(Action::State::FAILED);
569 get_model()->get_action_heap().remove(this);
570 cpu_->set_modified(true);
573 void CpuTiAction::suspend()
575 XBT_IN("(%p)", this);
577 set_suspend_state(Action::SuspendStates::SUSPENDED);
578 get_model()->get_action_heap().remove(this);
579 cpu_->set_modified(true);
584 void CpuTiAction::resume()
586 XBT_IN("(%p)", this);
587 if (is_suspended()) {
588 set_suspend_state(Action::SuspendStates::RUNNING);
589 cpu_->set_modified(true);
594 void CpuTiAction::set_sharing_penalty(double sharing_penalty)
596 XBT_IN("(%p,%g)", this, sharing_penalty);
597 set_sharing_penalty_no_update(sharing_penalty);
598 cpu_->set_modified(true);
602 double CpuTiAction::get_remains()
604 XBT_IN("(%p)", this);
605 cpu_->update_remaining_amount(EngineImpl::get_clock());
607 return get_remains_no_update();
610 } // namespace resource
611 } // namespace kernel
612 } // namespace simgrid