--- /dev/null
+/* Copyright (c) 2013-2023. The SimGrid Team. All rights reserved. */
+
+/* This program is free software; you can redistribute it and/or modify it
+ * under the terms of the license (GNU LGPL) which comes with this package. */
+
+#include "cpu_ti.hpp"
+#include "simgrid/kernel/routing/NetZoneImpl.hpp"
+#include "simgrid/s4u/Engine.hpp"
+#include "src/kernel/EngineImpl.hpp"
+#include "src/kernel/resource/profile/Event.hpp"
+#include "src/kernel/resource/profile/Profile.hpp"
+#include "xbt/asserts.h"
+
+#include <algorithm>
+#include <memory>
+
+constexpr double EPSILON = 0.000000001;
+
+XBT_LOG_NEW_DEFAULT_SUBCATEGORY(cpu_ti, res_cpu, "CPU resource, Trace Integration model");
+
+namespace simgrid::kernel::resource {
+
+/*********
+ * Trace *
+ *********/
+
+CpuTiProfile::CpuTiProfile(const profile::Profile* profile)
+{
+ double integral = 0;
+ double time = 0;
+ double prev_value = 1;
+ const std::vector<profile::DatedValue>& events = profile->get_event_list();
+ xbt_assert(not events.empty());
+ unsigned long nb_points = events.size() + 1;
+ time_points_.reserve(nb_points);
+ integral_.reserve(nb_points);
+ for (auto const& val : events) {
+ time += val.date_;
+ integral += val.date_ * prev_value;
+ time_points_.push_back(time);
+ integral_.push_back(integral);
+ prev_value = val.value_;
+ }
+
+ double delay = profile->get_repeat_delay() + events.at(0).date_;
+
+ xbt_assert(events.back().value_ == prev_value, "Profiles need to end as they start");
+ time += delay;
+ integral += delay * prev_value;
+
+ time_points_.push_back(time);
+ integral_.push_back(integral);
+}
+
+/**
+ * @brief Integrate trace
+ *
+ * Wrapper around surf_cpu_integrate_trace_simple() to get
+ * the cyclic effect.
+ *
+ * @param a Begin of interval
+ * @param b End of interval
+ * @return the integrate value. -1 if an error occurs.
+ */
+double CpuTiTmgr::integrate(double a, double b) const
+{
+ xbt_assert(a >= 0.0 && a <= b,
+ "Error, invalid integration interval [%.2f,%.2f]. You probably have a task executing with negative "
+ "computation amount. Check your code.",
+ a, b);
+ if (fabs(a - b) < EPSILON)
+ return 0.0;
+
+ if (type_ == Type::FIXED) {
+ return (b - a) * value_;
+ }
+
+ double a_index;
+ if (fabs(ceil(a / last_time_) - a / last_time_) < EPSILON)
+ a_index = 1 + ceil(a / last_time_);
+ else
+ a_index = ceil(a / last_time_);
+ double b_index = floor(b / last_time_);
+
+ if (a_index > b_index) { /* Same chunk */
+ return profile_->integrate_simple(a - (a_index - 1) * last_time_, b - b_index * last_time_);
+ }
+
+ double first_chunk = profile_->integrate_simple(a - (a_index - 1) * last_time_, last_time_);
+ double middle_chunk = (b_index - a_index) * total_;
+ double last_chunk = profile_->integrate_simple(0.0, b - b_index * last_time_);
+
+ XBT_DEBUG("first_chunk=%.2f middle_chunk=%.2f last_chunk=%.2f\n", first_chunk, middle_chunk, last_chunk);
+
+ return (first_chunk + middle_chunk + last_chunk);
+}
+
+/**
+ * @brief Auxiliary function to compute the integral between a and b.
+ * It simply computes the integrals at point a and b and returns the difference between them.
+ * @param a Initial point
+ * @param b Final point
+ */
+double CpuTiProfile::integrate_simple(double a, double b) const
+{
+ return integrate_simple_point(b) - integrate_simple_point(a);
+}
+
+/**
+ * @brief Auxiliary function to compute the integral at point a.
+ * @param a point
+ */
+double CpuTiProfile::integrate_simple_point(double a) const
+{
+ double integral = 0;
+ double a_aux = a;
+ long ind = binary_search(time_points_, a);
+ integral += integral_[ind];
+
+ XBT_DEBUG("a %f ind %ld integral %f ind + 1 %f ind %f time +1 %f time %f", a, ind, integral, integral_[ind + 1],
+ integral_[ind], time_points_[ind + 1], time_points_[ind]);
+ double_update(&a_aux, time_points_[ind], sg_precision_workamount * sg_precision_timing);
+ if (a_aux > 0)
+ integral +=
+ ((integral_[ind + 1] - integral_[ind]) / (time_points_[ind + 1] - time_points_[ind])) * (a - time_points_[ind]);
+ XBT_DEBUG("Integral a %f = %f", a, integral);
+
+ return integral;
+}
+
+/**
+ * @brief Computes the time needed to execute "amount" on cpu.
+ *
+ * Here, amount can span multiple trace periods
+ *
+ * @param a Initial time
+ * @param amount Amount to be executed
+ * @return End time
+ */
+double CpuTiTmgr::solve(double a, double amount) const
+{
+ /* Fix very small negative numbers */
+ if ((a < 0.0) && (a > -EPSILON)) {
+ a = 0.0;
+ }
+ if ((amount < 0.0) && (amount > -EPSILON)) {
+ amount = 0.0;
+ }
+
+ /* Sanity checks */
+ xbt_assert(a >= 0.0 && amount >= 0.0,
+ "Error, invalid parameters [a = %.2f, amount = %.2f]. "
+ "You probably have a task executing with negative computation amount. Check your code.",
+ a, amount);
+
+ /* At this point, a and amount are positive */
+ if (amount < EPSILON)
+ return a;
+
+ /* Is the trace fixed ? */
+ if (type_ == Type::FIXED) {
+ return (a + (amount / value_));
+ }
+
+ XBT_DEBUG("amount %f total %f", amount, total_);
+ /* Reduce the problem to one where amount <= trace_total */
+ double quotient = floor(amount / total_);
+ double reduced_amount = total_ * ((amount / total_) - floor(amount / total_));
+ double reduced_a = a - last_time_ * static_cast<int>(floor(a / last_time_));
+
+ XBT_DEBUG("Quotient: %g reduced_amount: %f reduced_a: %f", quotient, reduced_amount, reduced_a);
+
+ /* Now solve for new_amount which is <= trace_total */
+ XBT_DEBUG("Solve integral: [%.2f, amount=%.2f]", reduced_a, reduced_amount);
+
+ double amount_till_end = integrate(reduced_a, last_time_);
+ double reduced_b = amount_till_end > reduced_amount
+ ? profile_->solve_simple(reduced_a, reduced_amount)
+ : last_time_ + profile_->solve_simple(0.0, reduced_amount - amount_till_end);
+
+ /* Re-map to the original b and amount */
+ return last_time_ * floor(a / last_time_) + (quotient * last_time_) + reduced_b;
+}
+
+/**
+ * @brief Auxiliary function to solve integral.
+ * It returns the date when the requested amount of flops is available
+ * @param a Initial point
+ * @param amount Amount of flops
+ * @return The date when amount is available.
+ */
+double CpuTiProfile::solve_simple(double a, double amount) const
+{
+ double integral_a = integrate_simple_point(a);
+ long ind = binary_search(integral_, integral_a + amount);
+ double time = time_points_[ind];
+ time += (integral_a + amount - integral_[ind]) /
+ ((integral_[ind + 1] - integral_[ind]) / (time_points_[ind + 1] - time_points_[ind]));
+
+ return time;
+}
+
+/**
+ * @brief Auxiliary function to update the CPU speed scale.
+ *
+ * This function uses the trace structure to return the speed scale at the determined time a.
+ * @param a Time
+ * @return CPU speed scale
+ */
+double CpuTiTmgr::get_power_scale(double a) const
+{
+ double reduced_a = a - floor(a / last_time_) * last_time_;
+ long point = CpuTiProfile::binary_search(profile_->get_time_points(), reduced_a);
+ profile::DatedValue val = speed_profile_->get_event_list().at(point);
+ return val.value_;
+}
+
+/**
+ * @brief Creates a new integration trace from a tmgr_trace_t
+ *
+ * @param speed_trace CPU availability trace
+ * @param value Percentage of CPU speed available (useful to fixed tracing)
+ * @return Integration trace structure
+ */
+CpuTiTmgr::CpuTiTmgr(kernel::profile::Profile* speed_profile, double value) : speed_profile_(speed_profile)
+{
+ double total_time = 0.0;
+ profile_.reset(nullptr);
+
+ /* no availability file, fixed trace */
+ if (not speed_profile) {
+ value_ = value;
+ XBT_DEBUG("No availability trace. Constant value = %f", value);
+ return;
+ }
+
+ xbt_assert(speed_profile->is_repeating());
+
+ /* only one point available, fixed trace */
+ if (speed_profile->get_event_list().size() == 1) {
+ value_ = speed_profile->get_event_list().front().value_;
+ return;
+ }
+
+ type_ = Type::DYNAMIC;
+
+ /* count the total time of trace file */
+ for (auto const& val : speed_profile->get_event_list())
+ total_time += val.date_;
+ total_time += speed_profile->get_repeat_delay();
+
+ profile_ = std::make_unique<CpuTiProfile>(speed_profile);
+ last_time_ = total_time;
+ total_ = profile_->integrate_simple(0, total_time);
+
+ XBT_DEBUG("Total integral %f, last_time %f ", total_, last_time_);
+}
+
+/**
+ * @brief Binary search in array.
+ * It returns the last point of the interval in which "a" is.
+ * @param array Array
+ * @param a Value to search
+ * @return Index of point
+ */
+long CpuTiProfile::binary_search(const std::vector<double>& array, double a)
+{
+ if (array[0] > a)
+ return 0;
+ auto pos = std::upper_bound(begin(array), end(array), a);
+ return std::distance(begin(array), pos) - 1;
+}
+
+/*********
+ * Model *
+ *********/
+
+void CpuTiModel::create_pm_models()
+{
+ auto cpu_model_pm = std::make_shared<CpuTiModel>("Cpu_TI");
+ auto* engine = EngineImpl::get_instance();
+ engine->add_model(cpu_model_pm);
+ engine->get_netzone_root()->set_cpu_pm_model(cpu_model_pm);
+}
+
+CpuImpl* CpuTiModel::create_cpu(s4u::Host* host, const std::vector<double>& speed_per_pstate)
+{
+ return (new CpuTi(host, speed_per_pstate))->set_model(this);
+}
+
+double CpuTiModel::next_occurring_event(double now)
+{
+ double min_action_duration = -1;
+
+ /* iterates over modified cpus to update share resources */
+ for (auto it = std::begin(modified_cpus_); it != std::end(modified_cpus_);) {
+ CpuTi& cpu = *it;
+ ++it; // increment iterator here since the following call to ti.update_actions_finish_time() may invalidate it
+ cpu.update_actions_finish_time(now);
+ }
+
+ /* get the min next event if heap not empty */
+ if (not get_action_heap().empty())
+ min_action_duration = get_action_heap().top_date() - now;
+
+ XBT_DEBUG("Share resources, min next event date: %f", min_action_duration);
+
+ return min_action_duration;
+}
+
+void CpuTiModel::update_actions_state(double now, double /*delta*/)
+{
+ while (not get_action_heap().empty() && double_equals(get_action_heap().top_date(), now, sg_precision_timing)) {
+ auto* action = static_cast<CpuTiAction*>(get_action_heap().pop());
+ XBT_DEBUG("Action %p: finish", action);
+ action->finish(Action::State::FINISHED);
+ /* update remaining amount of all actions */
+ action->cpu_->update_remaining_amount(EngineImpl::get_clock());
+ }
+}
+
+/************
+ * Resource *
+ ************/
+CpuTi::CpuTi(s4u::Host* host, const std::vector<double>& speed_per_pstate) : CpuImpl(host, speed_per_pstate)
+{
+ speed_.peak = speed_per_pstate.front();
+ XBT_DEBUG("CPU create: peak=%f", speed_.peak);
+
+ speed_integrated_trace_ = new CpuTiTmgr(nullptr, 1 /*scale*/);
+}
+
+CpuTi::~CpuTi()
+{
+ set_modified(false);
+ delete speed_integrated_trace_;
+}
+
+CpuImpl* CpuTi::set_speed_profile(kernel::profile::Profile* profile)
+{
+ delete speed_integrated_trace_;
+ speed_integrated_trace_ = new CpuTiTmgr(profile, speed_.scale);
+
+ /* add a fake trace event if periodicity == 0 */
+ if (profile && profile->get_event_list().size() > 1) {
+ kernel::profile::DatedValue val = profile->get_event_list().back();
+ if (val.date_ < 1e-12) {
+ auto* prof = profile::ProfileBuilder::from_void();
+ speed_.event = prof->schedule(&profile::future_evt_set, this);
+ }
+ }
+ return this;
+}
+
+void CpuTi::apply_event(kernel::profile::Event* event, double value)
+{
+ if (event == speed_.event) {
+ XBT_DEBUG("Speed changed in trace! New fixed value: %f", value);
+
+ /* update remaining of actions and put in modified cpu list */
+ update_remaining_amount(EngineImpl::get_clock());
+
+ set_modified(true);
+
+ delete speed_integrated_trace_;
+ speed_integrated_trace_ = new CpuTiTmgr(value);
+
+ speed_.scale = value;
+ tmgr_trace_event_unref(&speed_.event);
+
+ } else if (event == get_state_event()) {
+ if (value > 0) {
+ if (not is_on()) {
+ XBT_VERB("Restart actors on host %s", get_iface()->get_cname());
+ get_iface()->turn_on();
+ }
+ } else {
+ get_iface()->turn_off();
+ double date = EngineImpl::get_clock();
+
+ /* put all action running on cpu to failed */
+ for (CpuTiAction& action : action_set_) {
+ if (action.get_state() == Action::State::INITED || action.get_state() == Action::State::STARTED ||
+ action.get_state() == Action::State::IGNORED) {
+ action.set_finish_time(date);
+ action.set_state(Action::State::FAILED);
+ get_model()->get_action_heap().remove(&action);
+ }
+ }
+ }
+ unref_state_event();
+
+ } else {
+ xbt_die("Unknown event!\n");
+ }
+}
+
+/** Update the actions that are running on this CPU (which was modified recently) */
+void CpuTi::update_actions_finish_time(double now)
+{
+ /* update remaining amount of actions */
+ update_remaining_amount(now);
+
+ /* Compute the sum of priorities for the actions running on that CPU */
+ sum_priority_ = 0.0;
+ for (CpuTiAction const& action : action_set_) {
+ /* action not running, skip it */
+ if (action.get_state_set() != get_model()->get_started_action_set())
+ continue;
+
+ /* bogus priority, skip it */
+ if (action.get_sharing_penalty() <= 0)
+ continue;
+
+ /* action suspended, skip it */
+ if (not action.is_running())
+ continue;
+
+ sum_priority_ += 1.0 / action.get_sharing_penalty();
+ }
+
+ for (CpuTiAction& action : action_set_) {
+ double min_finish = NO_MAX_DURATION;
+ /* action not running, skip it */
+ if (action.get_state_set() != get_model()->get_started_action_set())
+ continue;
+
+ /* verify if the action is really running on cpu */
+ if (action.is_running() && action.get_sharing_penalty() > 0) {
+ /* total area needed to finish the action. Used in trace integration */
+ double total_area = (action.get_remains() * sum_priority_ * action.get_sharing_penalty()) / speed_.peak;
+
+ action.set_finish_time(speed_integrated_trace_->solve(now, total_area));
+ /* verify which event will happen before (max_duration or finish time) */
+ if (action.get_max_duration() != NO_MAX_DURATION &&
+ action.get_start_time() + action.get_max_duration() < action.get_finish_time())
+ min_finish = action.get_start_time() + action.get_max_duration();
+ else
+ min_finish = action.get_finish_time();
+ } else {
+ /* put the max duration time on heap */
+ if (action.get_max_duration() != NO_MAX_DURATION)
+ min_finish = action.get_start_time() + action.get_max_duration();
+ }
+ /* add in action heap */
+ if (min_finish != NO_MAX_DURATION)
+ get_model()->get_action_heap().update(&action, min_finish, ActionHeap::Type::unset);
+ else
+ get_model()->get_action_heap().remove(&action);
+
+ XBT_DEBUG("Update finish time: Cpu(%s) Action: %p, Start Time: %f Finish Time: %f Max duration %f", get_cname(),
+ &action, action.get_start_time(), action.get_finish_time(), action.get_max_duration());
+ }
+ /* remove from modified cpu */
+ set_modified(false);
+}
+
+bool CpuTi::is_used() const
+{
+ return not action_set_.empty();
+}
+
+double CpuTi::get_speed_ratio()
+{
+ speed_.scale = speed_integrated_trace_->get_power_scale(EngineImpl::get_clock());
+ return CpuImpl::get_speed_ratio();
+}
+
+/** @brief Update the remaining amount of actions */
+void CpuTi::update_remaining_amount(double now)
+{
+ /* already up to date */
+ if (last_update_ >= now)
+ return;
+
+ /* compute the integration area */
+ double area_total = speed_integrated_trace_->integrate(last_update_, now) * speed_.peak;
+ XBT_DEBUG("Flops total: %f, Last update %f", area_total, last_update_);
+ for (CpuTiAction& action : action_set_) {
+ /* action not running, skip it */
+ if (action.get_state_set() != get_model()->get_started_action_set())
+ continue;
+
+ /* bogus priority, skip it */
+ if (action.get_sharing_penalty() <= 0)
+ continue;
+
+ /* action suspended, skip it */
+ if (not action.is_running())
+ continue;
+
+ /* action don't need update */
+ if (action.get_start_time() >= now)
+ continue;
+
+ /* skip action that are finishing now */
+ if (action.get_finish_time() >= 0 && action.get_finish_time() <= now)
+ continue;
+
+ /* update remaining */
+ action.update_remains(area_total / (sum_priority_ * action.get_sharing_penalty()));
+ XBT_DEBUG("Update remaining action(%p) remaining %f", &action, action.get_remains_no_update());
+ }
+ last_update_ = now;
+}
+
+CpuAction* CpuTi::execution_start(double size, double user_bound)
+{
+ XBT_IN("(%s,%g)", get_cname(), size);
+ xbt_assert(user_bound <= 0, "Invalid user bound (%lf) in CPU TI model", user_bound);
+ auto* action = new CpuTiAction(this, size);
+
+ action_set_.push_back(*action); // Actually start the action
+
+ XBT_OUT();
+ return action;
+}
+
+CpuAction* CpuTi::sleep(double duration)
+{
+ if (duration > 0)
+ duration = std::max(duration, sg_precision_timing);
+
+ XBT_IN("(%s,%g)", get_cname(), duration);
+ auto* action = new CpuTiAction(this, 1.0);
+
+ action->set_max_duration(duration);
+ action->set_suspend_state(Action::SuspendStates::SLEEPING);
+ if (duration == NO_MAX_DURATION)
+ action->set_state(Action::State::IGNORED);
+
+ action_set_.push_back(*action);
+
+ XBT_OUT();
+ return action;
+}
+
+void CpuTi::set_modified(bool modified)
+{
+ CpuTiList& modified_cpus = static_cast<CpuTiModel*>(get_model())->modified_cpus_;
+ if (modified) {
+ if (not cpu_ti_hook.is_linked()) {
+ modified_cpus.push_back(*this);
+ }
+ } else {
+ if (cpu_ti_hook.is_linked())
+ xbt::intrusive_erase(modified_cpus, *this);
+ }
+}
+
+/**********
+ * Action *
+ **********/
+
+CpuTiAction::CpuTiAction(CpuTi* cpu, double cost) : CpuAction(cpu->get_model(), cost, not cpu->is_on()), cpu_(cpu)
+{
+ cpu_->set_modified(true);
+}
+CpuTiAction::~CpuTiAction()
+{
+ /* remove from action_set */
+ if (action_ti_hook.is_linked())
+ xbt::intrusive_erase(cpu_->action_set_, *this);
+ /* remove from heap */
+ get_model()->get_action_heap().remove(this);
+ cpu_->set_modified(true);
+}
+
+void CpuTiAction::set_state(Action::State state)
+{
+ CpuAction::set_state(state);
+ cpu_->set_modified(true);
+}
+
+void CpuTiAction::cancel()
+{
+ this->set_state(Action::State::FAILED);
+ get_model()->get_action_heap().remove(this);
+ cpu_->set_modified(true);
+}
+
+void CpuTiAction::suspend()
+{
+ XBT_IN("(%p)", this);
+ if (is_running()) {
+ set_suspend_state(Action::SuspendStates::SUSPENDED);
+ get_model()->get_action_heap().remove(this);
+ cpu_->set_modified(true);
+ }
+ XBT_OUT();
+}
+
+void CpuTiAction::resume()
+{
+ XBT_IN("(%p)", this);
+ if (is_suspended()) {
+ set_suspend_state(Action::SuspendStates::RUNNING);
+ cpu_->set_modified(true);
+ }
+ XBT_OUT();
+}
+
+void CpuTiAction::set_sharing_penalty(double sharing_penalty)
+{
+ XBT_IN("(%p,%g)", this, sharing_penalty);
+ set_sharing_penalty_no_update(sharing_penalty);
+ cpu_->set_modified(true);
+ XBT_OUT();
+}
+
+double CpuTiAction::get_remains()
+{
+ XBT_IN("(%p)", this);
+ cpu_->update_remaining_amount(EngineImpl::get_clock());
+ XBT_OUT();
+ return get_remains_no_update();
+}
+
+} // namespace simgrid::kernel::resource
