1 /* Copyright (c) 2007-2021. 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 "ptask_L07.hpp"
7 #include "src/kernel/resource/profile/Event.hpp"
8 #include "surf/surf.hpp"
9 #include "xbt/config.hpp"
11 #include <unordered_set>
13 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(res_host);
14 XBT_LOG_EXTERNAL_CATEGORY(xbt_cfg);
16 /**************************************/
17 /*** Resource Creation & Destruction **/
18 /**************************************/
19 void surf_host_model_init_ptask_L07()
21 XBT_CINFO(xbt_cfg, "Switching to the L07 model to handle parallel tasks.");
23 surf_host_model = new simgrid::surf::HostL07Model();
24 all_existing_models.push_back(surf_host_model);
25 models_by_type[simgrid::kernel::resource::Model::Type::HOST].push_back(surf_host_model);
31 HostL07Model::HostL07Model() : HostModel()
33 auto* maxmin_system = new simgrid::kernel::lmm::FairBottleneck(true /* selective update */);
34 set_maxmin_system(maxmin_system);
35 network_model_ = std::make_unique<NetworkL07Model>(this, maxmin_system);
36 models_by_type[simgrid::kernel::resource::Model::Type::NETWORK].push_back(network_model_.get());
37 cpu_model_pm_ = std::make_unique<CpuL07Model>(this, maxmin_system);
38 models_by_type[simgrid::kernel::resource::Model::Type::CPU_PM].push_back(cpu_model_pm_.get());
41 HostL07Model::~HostL07Model() {}
43 CpuL07Model::CpuL07Model(HostL07Model* hmodel, kernel::lmm::System* sys)
44 : CpuModel(Model::UpdateAlgo::FULL), hostModel_(hmodel)
46 set_maxmin_system(sys);
49 CpuL07Model::~CpuL07Model()
51 set_maxmin_system(nullptr);
54 NetworkL07Model::NetworkL07Model(HostL07Model* hmodel, kernel::lmm::System* sys)
55 : NetworkModel(Model::UpdateAlgo::FULL), hostModel_(hmodel)
57 set_maxmin_system(sys);
58 loopback_ = NetworkL07Model::create_link(
59 "__loopback__", std::vector<double>{simgrid::config::get_value<double>("network/loopback-bw")},
60 s4u::Link::SharingPolicy::FATPIPE)
61 ->set_latency(simgrid::config::get_value<double>("network/loopback-lat"));
65 NetworkL07Model::~NetworkL07Model()
67 set_maxmin_system(nullptr);
70 double HostL07Model::next_occurring_event(double now)
72 double min = HostModel::next_occurring_event_full(now);
73 for (kernel::resource::Action const& action : *get_started_action_set()) {
74 const auto& net_action = static_cast<const L07Action&>(action);
75 if (net_action.get_latency() > 0 && (min < 0 || net_action.get_latency() < min)) {
76 min = net_action.get_latency();
77 XBT_DEBUG("Updating min with %p (start %f): %f", &net_action, net_action.get_start_time(), min);
80 XBT_DEBUG("min value: %f", min);
85 void HostL07Model::update_actions_state(double /*now*/, double delta)
87 for (auto it = std::begin(*get_started_action_set()); it != std::end(*get_started_action_set());) {
88 auto& action = static_cast<L07Action&>(*it);
89 ++it; // increment iterator here since the following calls to action.finish() may invalidate it
90 if (action.get_latency() > 0) {
91 if (action.get_latency() > delta) {
92 action.update_latency(delta, sg_surf_precision);
94 action.set_latency(0.0);
96 if ((action.get_latency() <= 0.0) && (action.is_suspended() == 0)) {
98 get_maxmin_system()->update_variable_penalty(action.get_variable(), 1.0);
99 action.set_last_update();
102 XBT_DEBUG("Action (%p) : remains (%g) updated by %g.", &action, action.get_remains(),
103 action.get_variable()->get_value() * delta);
104 action.update_remains(action.get_variable()->get_value() * delta);
105 action.update_max_duration(delta);
107 XBT_DEBUG("Action (%p) : remains (%g).", &action, action.get_remains());
109 /* In the next if cascade, the action can be finished either because:
110 * - The amount of remaining work reached 0
111 * - The max duration was reached
112 * If it's not done, it may have failed.
115 if (((action.get_remains() <= 0) && (action.get_variable()->get_penalty() > 0)) ||
116 ((action.get_max_duration() != NO_MAX_DURATION) && (action.get_max_duration() <= 0))) {
117 action.finish(kernel::resource::Action::State::FINISHED);
121 /* Need to check that none of the model has failed */
123 const kernel::lmm::Constraint* cnst = action.get_variable()->get_constraint(i);
124 while (cnst != nullptr) {
126 const kernel::resource::Resource* constraint_id = cnst->get_id();
127 if (not constraint_id->is_on()) {
128 XBT_DEBUG("Action (%p) Failed!!", &action);
129 action.finish(kernel::resource::Action::State::FAILED);
132 cnst = action.get_variable()->get_constraint(i);
137 kernel::resource::CpuAction* HostL07Model::execute_parallel(const std::vector<s4u::Host*>& host_list,
138 const double* flops_amount, const double* bytes_amount,
141 return new L07Action(this, host_list, flops_amount, bytes_amount, rate);
144 L07Action::L07Action(kernel::resource::Model* model, const std::vector<s4u::Host*>& host_list,
145 const double* flops_amount, const double* bytes_amount, double rate)
146 : CpuAction(model, 1.0, false), computationAmount_(flops_amount), communicationAmount_(bytes_amount), rate_(rate)
149 size_t used_host_nb = 0; /* Only the hosts with something to compute (>0 flops) are counted) */
150 double latency = 0.0;
151 this->set_last_update();
153 hostList_.insert(hostList_.end(), host_list.begin(), host_list.end());
155 if (flops_amount != nullptr)
156 used_host_nb += std::count_if(flops_amount, flops_amount + host_list.size(), [](double x) { return x > 0.0; });
158 /* Compute the number of affected resources... */
159 if (bytes_amount != nullptr) {
160 std::unordered_set<const char*> affected_links;
162 for (size_t k = 0; k < host_list.size() * host_list.size(); k++) {
163 if (bytes_amount[k] <= 0)
167 std::vector<kernel::resource::LinkImpl*> route;
168 hostList_[k / host_list.size()]->route_to(hostList_[k % host_list.size()], route, &lat);
169 latency = std::max(latency, lat);
171 for (auto const& link : route)
172 affected_links.insert(link->get_cname());
175 link_nb = affected_links.size();
178 XBT_DEBUG("Creating a parallel task (%p) with %zu hosts and %zu unique links.", this, host_list.size(), link_nb);
182 model->get_maxmin_system()->variable_new(this, 1.0, (rate > 0 ? rate : -1.0), host_list.size() + link_nb));
185 model->get_maxmin_system()->update_variable_penalty(get_variable(), 0.0);
187 /* Expand it for the CPUs even if there is nothing to compute, to make sure that it gets expended even if there is no
188 * communication either */
189 double bound = std::numeric_limits<double>::max();
190 for (size_t i = 0; i < host_list.size(); i++) {
191 model->get_maxmin_system()->expand(host_list[i]->pimpl_cpu->get_constraint(), get_variable(),
192 (flops_amount == nullptr ? 0.0 : flops_amount[i]));
193 if (flops_amount && flops_amount[i] > 0)
194 bound = std::min(bound, host_list[i]->pimpl_cpu->get_speed(1.0) * host_list[i]->pimpl_cpu->get_speed_ratio() /
197 if (bound < std::numeric_limits<double>::max())
198 model->get_maxmin_system()->update_variable_bound(get_variable(), bound);
200 if (bytes_amount != nullptr) {
201 for (size_t k = 0; k < host_list.size() * host_list.size(); k++) {
202 if (bytes_amount[k] <= 0.0)
204 std::vector<kernel::resource::LinkImpl*> route;
205 hostList_[k / host_list.size()]->route_to(hostList_[k % host_list.size()], route, nullptr);
207 for (auto const& link : route)
208 model->get_maxmin_system()->expand_add(link->get_constraint(), this->get_variable(), bytes_amount[k]);
212 if (link_nb + used_host_nb == 0) {
214 this->set_remains(0.0);
218 kernel::resource::Action* NetworkL07Model::communicate(s4u::Host* src, s4u::Host* dst, double size, double rate)
220 std::vector<s4u::Host*> host_list = {src, dst};
221 const auto* flops_amount = new double[2]();
222 auto* bytes_amount = new double[4]();
224 bytes_amount[1] = size;
226 kernel::resource::Action* res = hostModel_->execute_parallel(host_list, flops_amount, bytes_amount, rate);
227 static_cast<L07Action*>(res)->free_arrays_ = true;
231 kernel::resource::Cpu* CpuL07Model::create_cpu(s4u::Host* host, const std::vector<double>& speed_per_pstate)
233 return (new CpuL07(host, speed_per_pstate))->set_model(this);
236 kernel::resource::LinkImpl* NetworkL07Model::create_link(const std::string& name, const std::vector<double>& bandwidths,
237 s4u::Link::SharingPolicy policy)
239 xbt_assert(bandwidths.size() == 1, "Non WIFI link must have only 1 bandwidth.");
240 auto link = new LinkL07(name, bandwidths[0], policy, get_maxmin_system());
241 link->set_model(this);
249 kernel::resource::CpuAction* CpuL07::execution_start(double size)
251 std::vector<s4u::Host*> host_list = {get_iface()};
253 auto* flops_amount = new double[host_list.size()]();
254 flops_amount[0] = size;
256 kernel::resource::CpuAction* res =
257 static_cast<CpuL07Model*>(get_model())->hostModel_->execute_parallel(host_list, flops_amount, nullptr, -1);
258 static_cast<L07Action*>(res)->free_arrays_ = true;
262 kernel::resource::CpuAction* CpuL07::sleep(double duration)
264 auto* action = static_cast<L07Action*>(execution_start(1.0));
265 action->set_max_duration(duration);
266 action->set_suspend_state(kernel::resource::Action::SuspendStates::SLEEPING);
267 get_model()->get_maxmin_system()->update_variable_penalty(action->get_variable(), 0.0);
272 bool CpuL07::is_used() const
274 return get_model()->get_maxmin_system()->constraint_used(get_constraint());
277 /** @brief take into account changes of speed (either load or max) */
278 void CpuL07::on_speed_change()
280 const kernel::lmm::Variable* var;
281 const kernel::lmm::Element* elem = nullptr;
283 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), speed_.peak * speed_.scale);
284 while ((var = get_constraint()->get_variable(&elem))) {
285 const kernel::resource::Action* action = var->get_id();
287 get_model()->get_maxmin_system()->update_variable_bound(action->get_variable(), speed_.scale * speed_.peak);
290 Cpu::on_speed_change();
293 LinkL07::LinkL07(const std::string& name, double bandwidth, s4u::Link::SharingPolicy policy,
294 kernel::lmm::System* system)
297 this->set_constraint(system->constraint_new(this, bandwidth));
298 bandwidth_.peak = bandwidth;
300 if (policy == s4u::Link::SharingPolicy::FATPIPE)
301 get_constraint()->unshare();
304 bool LinkL07::is_used() const
306 return get_model()->get_maxmin_system()->constraint_used(get_constraint());
309 void CpuL07::apply_event(kernel::profile::Event* triggered, double value)
311 XBT_DEBUG("Updating cpu %s (%p) with value %g", get_cname(), this, value);
312 if (triggered == speed_.event) {
313 speed_.scale = value;
315 tmgr_trace_event_unref(&speed_.event);
317 } else if (triggered == state_event_) {
320 XBT_VERB("Restart actors on host %s", get_iface()->get_cname());
321 get_iface()->turn_on();
324 get_iface()->turn_off();
325 tmgr_trace_event_unref(&state_event_);
328 xbt_die("Unknown event!\n");
332 void LinkL07::apply_event(kernel::profile::Event* triggered, double value)
334 XBT_DEBUG("Updating link %s (%p) with value=%f", get_cname(), this, value);
335 if (triggered == bandwidth_.event) {
336 set_bandwidth(value);
337 tmgr_trace_event_unref(&bandwidth_.event);
339 } else if (triggered == latency_.event) {
341 tmgr_trace_event_unref(&latency_.event);
343 } else if (triggered == state_event_) {
348 tmgr_trace_event_unref(&state_event_);
351 xbt_die("Unknown event ! \n");
355 void LinkL07::set_bandwidth(double value)
357 bandwidth_.peak = value;
358 LinkImpl::on_bandwidth_change();
360 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), bandwidth_.peak * bandwidth_.scale);
363 kernel::resource::LinkImpl* LinkL07::set_latency(double value)
365 latency_check(value);
366 const kernel::lmm::Variable* var;
368 const kernel::lmm::Element* elem = nullptr;
370 latency_.peak = value;
371 while ((var = get_constraint()->get_variable(&elem))) {
372 action = static_cast<L07Action*>(var->get_id());
373 action->updateBound();
377 LinkL07::~LinkL07() = default;
383 L07Action::~L07Action()
386 delete[] computationAmount_;
387 delete[] communicationAmount_;
391 void L07Action::updateBound()
393 double lat_current = 0.0;
395 size_t host_count = hostList_.size();
397 if (communicationAmount_ != nullptr) {
398 for (size_t i = 0; i < host_count; i++) {
399 for (size_t j = 0; j < host_count; j++) {
400 if (communicationAmount_[i * host_count + j] > 0) {
402 std::vector<kernel::resource::LinkImpl*> route;
403 hostList_.at(i)->route_to(hostList_.at(j), route, &lat);
405 lat_current = std::max(lat_current, lat * communicationAmount_[i * host_count + j]);
410 double lat_bound = kernel::resource::NetworkModel::cfg_tcp_gamma / (2.0 * lat_current);
411 XBT_DEBUG("action (%p) : lat_bound = %g", this, lat_bound);
412 if ((latency_ <= 0.0) && is_running()) {
414 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), lat_bound);
416 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), std::min(rate_, lat_bound));
421 } // namespace simgrid