1 /* Copyright (c) 2007-2022. 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. */
6 #include <simgrid/kernel/routing/NetZoneImpl.hpp>
7 #include <simgrid/s4u/Engine.hpp>
8 #include <xbt/config.hpp>
10 #include "src/kernel/EngineImpl.hpp"
11 #include "src/kernel/resource/profile/Event.hpp"
12 #include "src/surf/ptask_L07.hpp"
14 #include <unordered_set>
16 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(res_host);
17 XBT_LOG_EXTERNAL_CATEGORY(xbt_cfg);
19 /**************************************/
20 /*** Resource Creation & Destruction **/
21 /**************************************/
22 void surf_host_model_init_ptask_L07()
24 XBT_CINFO(xbt_cfg, "Switching to the L07 model to handle parallel tasks.");
26 auto host_model = std::make_shared<simgrid::kernel::resource::HostL07Model>("Host_Ptask");
27 auto* engine = simgrid::kernel::EngineImpl::get_instance();
28 engine->add_model(host_model);
29 engine->get_netzone_root()->set_host_model(host_model);
36 HostL07Model::HostL07Model(const std::string& name) : HostModel(name)
38 auto* maxmin_system = new lmm::FairBottleneck(true /* selective update */);
39 set_maxmin_system(maxmin_system);
41 auto net_model = std::make_shared<NetworkL07Model>("Network_Ptask", this, maxmin_system);
42 auto engine = EngineImpl::get_instance();
43 engine->add_model(net_model);
44 engine->get_netzone_root()->set_network_model(net_model);
46 auto cpu_model = std::make_shared<CpuL07Model>("Cpu_Ptask", this, maxmin_system);
47 engine->add_model(cpu_model);
48 engine->get_netzone_root()->set_cpu_pm_model(cpu_model);
51 CpuL07Model::CpuL07Model(const std::string& name, HostL07Model* hmodel, lmm::System* sys)
52 : CpuModel(name), hostModel_(hmodel)
54 set_maxmin_system(sys);
57 CpuL07Model::~CpuL07Model()
59 set_maxmin_system(nullptr);
62 NetworkL07Model::NetworkL07Model(const std::string& name, HostL07Model* hmodel, lmm::System* sys)
63 : NetworkModel(name), hostModel_(hmodel)
65 set_maxmin_system(sys);
66 loopback_ = create_link("__loopback__", {simgrid::config::get_value<double>("network/loopback-bw")});
67 loopback_->set_sharing_policy(s4u::Link::SharingPolicy::FATPIPE, {});
68 loopback_->set_latency(simgrid::config::get_value<double>("network/loopback-lat"));
72 NetworkL07Model::~NetworkL07Model()
74 set_maxmin_system(nullptr);
77 double HostL07Model::next_occurring_event(double now)
79 double min = HostModel::next_occurring_event_full(now);
80 for (Action const& action : *get_started_action_set()) {
81 const auto& net_action = static_cast<const L07Action&>(action);
82 if (net_action.get_latency() > 0 && (min < 0 || net_action.get_latency() < min)) {
83 min = net_action.get_latency();
84 XBT_DEBUG("Updating min with %p (start %f): %f", &net_action, net_action.get_start_time(), min);
87 XBT_DEBUG("min value: %f", min);
92 void HostL07Model::update_actions_state(double /*now*/, double delta)
94 for (auto it = std::begin(*get_started_action_set()); it != std::end(*get_started_action_set());) {
95 auto& action = static_cast<L07Action&>(*it);
96 ++it; // increment iterator here since the following calls to action.finish() may invalidate it
97 if (action.get_latency() > 0) {
98 if (action.get_latency() > delta) {
99 action.update_latency(delta, sg_surf_precision);
101 action.set_latency(0.0);
103 if ((action.get_latency() <= 0.0) && (action.is_suspended() == 0)) {
104 action.updateBound();
105 get_maxmin_system()->update_variable_penalty(action.get_variable(), 1.0);
106 action.set_last_update();
109 XBT_DEBUG("Action (%p) : remains (%g) updated by %g.", &action, action.get_remains(), action.get_rate() * delta);
110 action.update_remains(action.get_rate() * delta);
111 action.update_max_duration(delta);
113 XBT_DEBUG("Action (%p) : remains (%g).", &action, action.get_remains());
115 /* In the next if cascade, the action can be finished either because:
116 * - The amount of remaining work reached 0
117 * - The max duration was reached
118 * If it's not done, it may have failed.
121 if (((action.get_remains() <= 0) && (action.get_variable()->get_penalty() > 0)) ||
122 ((action.get_max_duration() != NO_MAX_DURATION) && (action.get_max_duration() <= 0))) {
123 action.finish(Action::State::FINISHED);
127 /* Need to check that none of the model has failed */
129 const lmm::Constraint* cnst = action.get_variable()->get_constraint(i);
130 while (cnst != nullptr) {
132 const Resource* constraint_id = cnst->get_id();
133 if (not constraint_id->is_on()) {
134 XBT_DEBUG("Action (%p) Failed!!", &action);
135 action.finish(Action::State::FAILED);
138 cnst = action.get_variable()->get_constraint(i);
143 CpuAction* HostL07Model::execute_parallel(const std::vector<s4u::Host*>& host_list, const double* flops_amount,
144 const double* bytes_amount, double rate)
146 return new L07Action(this, host_list, flops_amount, bytes_amount, rate);
149 L07Action::L07Action(Model* model, const std::vector<s4u::Host*>& host_list, const double* flops_amount,
150 const double* bytes_amount, double rate)
151 : CpuAction(model, 1.0, false), computationAmount_(flops_amount), communicationAmount_(bytes_amount), rate_(rate)
154 size_t used_host_nb = 0; /* Only the hosts with something to compute (>0 flops) are counted) */
155 double latency = 0.0;
156 this->set_last_update();
158 hostList_.insert(hostList_.end(), host_list.begin(), host_list.end());
160 if (flops_amount != nullptr)
161 used_host_nb += std::count_if(flops_amount, flops_amount + host_list.size(), [](double x) { return x > 0.0; });
163 /* Compute the number of affected resources... */
164 if (bytes_amount != nullptr) {
165 std::unordered_set<const char*> affected_links;
167 for (size_t k = 0; k < host_list.size() * host_list.size(); k++) {
168 if (bytes_amount[k] <= 0)
172 std::vector<StandardLinkImpl*> route;
173 hostList_[k / host_list.size()]->route_to(hostList_[k % host_list.size()], route, &lat);
174 latency = std::max(latency, lat);
176 for (auto const& link : route)
177 affected_links.insert(link->get_cname());
180 link_nb = affected_links.size();
183 XBT_DEBUG("Creating a parallel task (%p) with %zu hosts and %zu unique links.", this, host_list.size(), link_nb);
187 model->get_maxmin_system()->variable_new(this, 1.0, (rate > 0 ? rate : -1.0), host_list.size() + link_nb));
190 model->get_maxmin_system()->update_variable_penalty(get_variable(), 0.0);
192 /* Expand it for the CPUs even if there is nothing to compute, to make sure that it gets expended even if there is no
193 * communication either */
194 for (size_t i = 0; i < host_list.size(); i++) {
195 model->get_maxmin_system()->expand(host_list[i]->get_cpu()->get_constraint(), get_variable(),
196 (flops_amount == nullptr ? 0.0 : flops_amount[i]));
199 if (bytes_amount != nullptr) {
200 for (size_t k = 0; k < host_list.size() * host_list.size(); k++) {
201 if (bytes_amount[k] <= 0.0)
203 std::vector<StandardLinkImpl*> route;
204 hostList_[k / host_list.size()]->route_to(hostList_[k % host_list.size()], route, nullptr);
206 for (auto const& link : route)
207 model->get_maxmin_system()->expand_add(link->get_constraint(), this->get_variable(), bytes_amount[k]);
211 if (link_nb + used_host_nb == 0) {
213 this->set_remains(0.0);
215 /* finally calculate the initial bound value */
219 Action* NetworkL07Model::communicate(s4u::Host* src, s4u::Host* dst, double size, double rate)
221 std::vector<s4u::Host*> host_list = {src, dst};
222 const auto* flops_amount = new double[2]();
223 auto* bytes_amount = new double[4]();
225 bytes_amount[1] = size;
227 Action* res = hostModel_->execute_parallel(host_list, flops_amount, bytes_amount, rate);
228 static_cast<L07Action*>(res)->free_arrays_ = true;
232 CpuImpl* CpuL07Model::create_cpu(s4u::Host* host, const std::vector<double>& speed_per_pstate)
234 return (new CpuL07(host, speed_per_pstate))->set_model(this);
237 StandardLinkImpl* NetworkL07Model::create_link(const std::string& name, const std::vector<double>& bandwidths)
239 xbt_assert(bandwidths.size() == 1, "Non WIFI link must have only 1 bandwidth.");
240 auto link = new LinkL07(name, bandwidths[0], get_maxmin_system());
241 link->set_model(this);
245 StandardLinkImpl* NetworkL07Model::create_wifi_link(const std::string& name, const std::vector<double>& bandwidths)
254 CpuAction* CpuL07::execution_start(double size, double user_bound)
256 std::vector<s4u::Host*> host_list = {get_iface()};
257 xbt_assert(user_bound <= 0, "User bound not supported by ptask model");
259 auto* flops_amount = new double[host_list.size()]();
260 flops_amount[0] = size;
263 static_cast<CpuL07Model*>(get_model())->hostModel_->execute_parallel(host_list, flops_amount, nullptr, -1);
264 static_cast<L07Action*>(res)->free_arrays_ = true;
268 CpuAction* CpuL07::sleep(double duration)
270 auto* action = static_cast<L07Action*>(execution_start(1.0, -1));
271 action->set_max_duration(duration);
272 action->set_suspend_state(Action::SuspendStates::SLEEPING);
273 get_model()->get_maxmin_system()->update_variable_penalty(action->get_variable(), 0.0);
278 /** @brief take into account changes of speed (either load or max) */
279 void CpuL07::on_speed_change()
281 const lmm::Element* elem = nullptr;
283 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), get_core_count() * speed_.peak * speed_.scale);
285 while (const auto* var = get_constraint()->get_variable(&elem)) {
286 auto* action = static_cast<L07Action*>(var->get_id());
287 action->updateBound();
290 CpuImpl::on_speed_change();
293 LinkL07::LinkL07(const std::string& name, double bandwidth, lmm::System* system) : StandardLinkImpl(name)
295 this->set_constraint(system->constraint_new(this, bandwidth));
296 bandwidth_.peak = bandwidth;
299 void CpuL07::apply_event(profile::Event* triggered, double value)
301 XBT_DEBUG("Updating cpu %s (%p) with value %g", get_cname(), this, value);
302 if (triggered == speed_.event) {
303 speed_.scale = value;
305 tmgr_trace_event_unref(&speed_.event);
307 } else if (triggered == get_state_event()) {
310 XBT_VERB("Restart actors on host %s", get_iface()->get_cname());
311 get_iface()->turn_on();
314 get_iface()->turn_off();
318 xbt_die("Unknown event!\n");
322 void LinkL07::apply_event(profile::Event* triggered, double value)
324 XBT_DEBUG("Updating link %s (%p) with value=%f", get_cname(), this, value);
325 if (triggered == bandwidth_.event) {
326 set_bandwidth(value);
327 tmgr_trace_event_unref(&bandwidth_.event);
329 } else if (triggered == latency_.event) {
331 tmgr_trace_event_unref(&latency_.event);
333 } else if (triggered == get_state_event()) {
340 xbt_die("Unknown event ! \n");
344 void LinkL07::set_bandwidth(double value)
346 bandwidth_.peak = value;
347 StandardLinkImpl::on_bandwidth_change();
349 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), bandwidth_.peak * bandwidth_.scale);
352 void LinkL07::set_latency(double value)
354 latency_check(value);
355 const lmm::Element* elem = nullptr;
357 latency_.peak = value;
358 while (const auto* var = get_constraint()->get_variable(&elem)) {
359 auto* action = static_cast<L07Action*>(var->get_id());
360 action->updateBound();
363 LinkL07::~LinkL07() = default;
369 L07Action::~L07Action()
372 delete[] computationAmount_;
373 delete[] communicationAmount_;
377 double L07Action::calculateNetworkBound()
379 double lat_current = 0.0;
380 double lat_bound = std::numeric_limits<double>::max();
382 size_t host_count = hostList_.size();
384 if (communicationAmount_ == nullptr) {
388 for (size_t i = 0; i < host_count; i++) {
389 for (size_t j = 0; j < host_count; j++) {
390 if (communicationAmount_[i * host_count + j] > 0) {
392 std::vector<StandardLinkImpl*> route;
393 hostList_.at(i)->route_to(hostList_.at(j), route, &lat);
395 lat_current = std::max(lat_current, lat * communicationAmount_[i * host_count + j]);
399 if (lat_current > 0) {
400 lat_bound = NetworkModel::cfg_tcp_gamma / (2.0 * lat_current);
405 double L07Action::calculateCpuBound()
407 double cpu_bound = std::numeric_limits<double>::max();
409 if (computationAmount_ == nullptr) {
413 for (size_t i = 0; i < hostList_.size(); i++) {
414 if (computationAmount_[i] > 0) {
415 cpu_bound = std::min(cpu_bound, hostList_[i]->get_cpu()->get_speed(1.0) *
416 hostList_[i]->get_cpu()->get_speed_ratio() / computationAmount_[i]);
422 void L07Action::updateBound()
424 double bound = std::min(calculateNetworkBound(), calculateCpuBound());
426 XBT_DEBUG("action (%p) : bound = %g", this, bound);
428 /* latency has been paid (or no latency), we can set the appropriate bound for multicore or network limit */
429 if ((bound < std::numeric_limits<double>::max()) && (latency_ <= 0.0)) {
431 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), bound);
433 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), std::min(rate_, bound));
437 } // namespace resource
438 } // namespace kernel
439 } // namespace simgrid