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 "simgrid/config.h"
11 #include "src/kernel/EngineImpl.hpp"
12 #if SIMGRID_HAVE_EIGEN3
13 #include "src/kernel/lmm/bmf.hpp"
15 #include "src/kernel/resource/profile/Event.hpp"
16 #include "src/surf/ptask_L07.hpp"
18 #include <unordered_set>
20 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(res_host);
21 XBT_LOG_EXTERNAL_CATEGORY(xbt_cfg);
26 static simgrid::config::Flag<std::string> cfg_ptask_solver("host/solver",
27 "Set linear equations solver used by ptask model",
29 &simgrid::kernel::lmm::System::validate_solver);
31 /**************************************/
32 /*** Resource Creation & Destruction **/
33 /**************************************/
34 void surf_host_model_init_ptask_L07()
36 XBT_CINFO(xbt_cfg, "Switching to the L07 model to handle parallel tasks.");
37 xbt_assert(cfg_ptask_solver != "maxmin", "Invalid configuration. Cannot use maxmin solver with parallel tasks.");
39 auto* system = simgrid::kernel::lmm::System::build(cfg_ptask_solver, true /* selective update */);
40 auto host_model = std::make_shared<simgrid::kernel::resource::HostL07Model>("Host_Ptask", system);
41 auto* engine = simgrid::kernel::EngineImpl::get_instance();
42 engine->add_model(host_model);
43 engine->get_netzone_root()->set_host_model(host_model);
46 namespace simgrid::kernel::resource {
48 HostL07Model::HostL07Model(const std::string& name, lmm::System* sys) : HostModel(name)
50 set_maxmin_system(sys);
52 auto net_model = std::make_shared<NetworkL07Model>("Network_Ptask", this, sys);
53 auto engine = EngineImpl::get_instance();
54 engine->add_model(net_model);
55 engine->get_netzone_root()->set_network_model(net_model);
57 auto cpu_model = std::make_shared<CpuL07Model>("Cpu_Ptask", this, sys);
58 engine->add_model(cpu_model);
59 engine->get_netzone_root()->set_cpu_pm_model(cpu_model);
62 CpuL07Model::CpuL07Model(const std::string& name, HostL07Model* hmodel, lmm::System* sys)
63 : CpuModel(name), hostModel_(hmodel)
65 set_maxmin_system(sys);
68 CpuL07Model::~CpuL07Model()
70 set_maxmin_system(nullptr);
73 NetworkL07Model::NetworkL07Model(const std::string& name, HostL07Model* hmodel, lmm::System* sys)
74 : NetworkModel(name), hostModel_(hmodel)
76 set_maxmin_system(sys);
77 loopback_.reset(create_link("__loopback__", {simgrid::config::get_value<double>("network/loopback-bw")}));
78 loopback_->set_sharing_policy(s4u::Link::SharingPolicy::FATPIPE, {});
79 loopback_->set_latency(simgrid::config::get_value<double>("network/loopback-lat"));
80 loopback_->get_iface()->seal();
83 NetworkL07Model::~NetworkL07Model()
85 set_maxmin_system(nullptr);
88 double HostL07Model::next_occurring_event(double now)
90 double min = HostModel::next_occurring_event_full(now);
91 for (Action const& action : *get_started_action_set()) {
92 const auto& net_action = static_cast<const L07Action&>(action);
93 if (net_action.get_latency() > 0 && (min < 0 || net_action.get_latency() < min)) {
94 min = net_action.get_latency();
95 XBT_DEBUG("Updating min with %p (start %f): %f", &net_action, net_action.get_start_time(), min);
98 XBT_DEBUG("min value: %f", min);
103 void HostL07Model::update_actions_state(double /*now*/, double delta)
105 for (auto it = std::begin(*get_started_action_set()); it != std::end(*get_started_action_set());) {
106 auto& action = static_cast<L07Action&>(*it);
107 ++it; // increment iterator here since the following calls to action.finish() may invalidate it
108 if (action.get_latency() > 0) {
109 if (action.get_latency() > delta) {
110 action.update_latency(delta, sg_surf_precision);
112 action.set_latency(0.0);
114 if ((action.get_latency() <= 0.0) && (action.is_suspended() == 0)) {
115 action.update_bound();
116 get_maxmin_system()->update_variable_penalty(action.get_variable(), 1.0);
117 action.set_last_update();
120 XBT_DEBUG("Action (%p) : remains (%g) updated by %g.", &action, action.get_remains(), action.get_rate() * delta);
121 action.update_remains(action.get_rate() * delta);
122 action.update_max_duration(delta);
124 XBT_DEBUG("Action (%p) : remains (%g).", &action, action.get_remains());
126 /* In the next if cascade, the action can be finished either because:
127 * - The amount of remaining work reached 0
128 * - The max duration was reached
129 * If it's not done, it may have failed.
132 if (((action.get_remains() <= 0) && (action.get_variable()->get_penalty() > 0)) ||
133 ((action.get_max_duration() != NO_MAX_DURATION) && (action.get_max_duration() <= 0))) {
134 action.finish(Action::State::FINISHED);
138 /* Need to check that none of the model has failed */
140 const lmm::Constraint* cnst = action.get_variable()->get_constraint(i);
141 while (cnst != nullptr) {
143 if (not cnst->get_id()->is_on()) {
144 XBT_DEBUG("Action (%p) Failed!!", &action);
145 action.finish(Action::State::FAILED);
148 cnst = action.get_variable()->get_constraint(i);
153 CpuAction* HostL07Model::execute_parallel(const std::vector<s4u::Host*>& host_list, const double* flops_amount,
154 const double* bytes_amount, double rate)
156 return new L07Action(this, host_list, flops_amount, bytes_amount, rate);
159 L07Action::L07Action(Model* model, const std::vector<s4u::Host*>& host_list, const double* flops_amount,
160 const double* bytes_amount, double rate)
161 : CpuAction(model, 1.0, false)
162 , host_list_(host_list)
163 , computation_amount_(flops_amount)
164 , communication_amount_(bytes_amount)
168 const size_t host_nb = host_list_.size();
169 size_t used_host_nb = 0; /* Only the hosts with something to compute (>0 flops) are counted) */
170 double latency = 0.0;
171 this->set_last_update();
173 if (flops_amount != nullptr)
174 used_host_nb += std::count_if(flops_amount, flops_amount + host_nb, [](double x) { return x > 0.0; });
176 /* Compute the number of affected resources... */
177 if (bytes_amount != nullptr) {
178 std::unordered_set<const char*> affected_links;
180 for (size_t k = 0; k < host_nb * host_nb; k++) {
181 if (bytes_amount[k] <= 0)
185 std::vector<StandardLinkImpl*> route;
186 host_list_[k / host_nb]->route_to(host_list_[k % host_nb], route, &lat);
187 latency = std::max(latency, lat);
189 for (auto const& link : route)
190 affected_links.insert(link->get_cname());
193 link_nb = affected_links.size();
196 XBT_DEBUG("Creating a parallel task (%p) with %zu hosts and %zu unique links.", this, host_nb, link_nb);
199 set_variable(model->get_maxmin_system()->variable_new(this, 1.0, (rate > 0 ? rate : -1.0), host_nb + link_nb));
202 model->get_maxmin_system()->update_variable_penalty(get_variable(), 0.0);
204 /* Expand it for the CPUs even if there is nothing to compute, to make sure that it gets expended even if there is no
205 * communication either */
206 for (size_t i = 0; i < host_nb; i++) {
207 model->get_maxmin_system()->expand(host_list[i]->get_cpu()->get_constraint(), get_variable(),
208 (flops_amount == nullptr ? 0.0 : flops_amount[i]), true);
211 if (bytes_amount != nullptr) {
212 for (size_t k = 0; k < host_nb * host_nb; k++) {
213 if (bytes_amount[k] <= 0.0)
215 std::vector<StandardLinkImpl*> route;
216 host_list_[k / host_nb]->route_to(host_list_[k % host_nb], route, nullptr);
218 for (auto const& link : route)
219 model->get_maxmin_system()->expand(link->get_constraint(), this->get_variable(), bytes_amount[k]);
223 if (link_nb + used_host_nb == 0) {
225 this->set_remains(0.0);
227 /* finally calculate the initial bound value */
231 Action* NetworkL07Model::communicate(s4u::Host* src, s4u::Host* dst, double size, double rate)
233 std::vector<s4u::Host*> host_list = {src, dst};
234 const auto* flops_amount = new double[2]();
235 auto* bytes_amount = new double[4]();
237 bytes_amount[1] = size;
239 Action* res = hostModel_->execute_parallel(host_list, flops_amount, bytes_amount, rate);
240 static_cast<L07Action*>(res)->free_arrays_ = true;
244 CpuImpl* CpuL07Model::create_cpu(s4u::Host* host, const std::vector<double>& speed_per_pstate)
246 return (new CpuL07(host, speed_per_pstate))->set_model(this);
249 StandardLinkImpl* NetworkL07Model::create_link(const std::string& name, const std::vector<double>& bandwidths)
251 xbt_assert(bandwidths.size() == 1, "Non WIFI link must have only 1 bandwidth.");
252 auto link = new LinkL07(name, bandwidths[0], get_maxmin_system());
253 link->set_model(this);
257 StandardLinkImpl* NetworkL07Model::create_wifi_link(const std::string& name, const std::vector<double>& bandwidths)
266 CpuAction* CpuL07::execution_start(double size, double user_bound)
268 std::vector<s4u::Host*> host_list = {get_iface()};
269 xbt_assert(user_bound <= 0, "User bound not supported by ptask model");
271 auto* flops_amount = new double[host_list.size()]();
272 flops_amount[0] = size;
275 static_cast<CpuL07Model*>(get_model())->hostModel_->execute_parallel(host_list, flops_amount, nullptr, -1);
276 static_cast<L07Action*>(res)->free_arrays_ = true;
280 CpuAction* CpuL07::sleep(double duration)
282 auto* action = static_cast<L07Action*>(execution_start(1.0, -1));
283 action->set_max_duration(duration);
284 action->set_suspend_state(Action::SuspendStates::SLEEPING);
285 get_model()->get_maxmin_system()->update_variable_penalty(action->get_variable(), 0.0);
290 /** @brief take into account changes of speed (either load or max) */
291 void CpuL07::on_speed_change()
293 const lmm::Element* elem = nullptr;
295 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), get_core_count() * speed_.peak * speed_.scale);
297 while (const auto* var = get_constraint()->get_variable(&elem)) {
298 const auto* action = static_cast<L07Action*>(var->get_id());
299 action->update_bound();
302 CpuImpl::on_speed_change();
305 LinkL07::LinkL07(const std::string& name, double bandwidth, lmm::System* system) : StandardLinkImpl(name)
307 this->set_constraint(system->constraint_new(this, bandwidth));
308 bandwidth_.peak = bandwidth;
311 void CpuL07::apply_event(profile::Event* triggered, double value)
313 XBT_DEBUG("Updating cpu %s (%p) with value %g", get_cname(), this, value);
314 if (triggered == speed_.event) {
315 speed_.scale = value;
317 tmgr_trace_event_unref(&speed_.event);
319 } else if (triggered == get_state_event()) {
322 XBT_VERB("Restart actors on host %s", get_iface()->get_cname());
323 get_iface()->turn_on();
326 get_iface()->turn_off();
330 xbt_die("Unknown event!\n");
334 void LinkL07::apply_event(profile::Event* triggered, double value)
336 XBT_DEBUG("Updating link %s (%p) with value=%f", get_cname(), this, value);
337 if (triggered == bandwidth_.event) {
338 set_bandwidth(value);
339 tmgr_trace_event_unref(&bandwidth_.event);
341 } else if (triggered == latency_.event) {
343 tmgr_trace_event_unref(&latency_.event);
345 } else if (triggered == get_state_event()) {
352 xbt_die("Unknown event ! \n");
356 void LinkL07::set_bandwidth(double value)
358 bandwidth_.peak = value;
359 StandardLinkImpl::on_bandwidth_change();
361 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), bandwidth_.peak * bandwidth_.scale);
364 void LinkL07::set_latency(double value)
366 latency_check(value);
367 const lmm::Element* elem = nullptr;
369 latency_.peak = value;
370 while (const auto* var = get_constraint()->get_variable(&elem)) {
371 const auto* action = static_cast<L07Action*>(var->get_id());
372 action->update_bound();
375 LinkL07::~LinkL07() = default;
381 L07Action::~L07Action()
384 delete[] computation_amount_;
385 delete[] communication_amount_;
389 double L07Action::calculate_network_bound() const
391 double lat_current = 0.0;
392 double lat_bound = std::numeric_limits<double>::max();
394 size_t host_count = host_list_.size();
396 if (communication_amount_ == nullptr) {
400 for (size_t i = 0; i < host_count; i++) {
401 for (size_t j = 0; j < host_count; j++) {
402 if (communication_amount_[i * host_count + j] > 0) {
404 std::vector<StandardLinkImpl*> route;
405 host_list_.at(i)->route_to(host_list_.at(j), route, &lat);
407 lat_current = std::max(lat_current, lat * communication_amount_[i * host_count + j]);
411 if (lat_current > 0) {
412 lat_bound = NetworkModel::cfg_tcp_gamma / (2.0 * lat_current);
417 double L07Action::calculate_cpu_bound() const
419 double cpu_bound = std::numeric_limits<double>::max();
421 if (computation_amount_ == nullptr) {
425 for (size_t i = 0; i < host_list_.size(); i++) {
426 if (computation_amount_[i] > 0) {
427 cpu_bound = std::min(cpu_bound, host_list_[i]->get_cpu()->get_speed(1.0) *
428 host_list_[i]->get_cpu()->get_speed_ratio() / computation_amount_[i]);
434 void L07Action::update_bound() const
436 double bound = std::min(calculate_network_bound(), calculate_cpu_bound());
438 XBT_DEBUG("action (%p) : bound = %g", this, bound);
440 /* latency has been paid (or no latency), we can set the appropriate bound for multicore or network limit */
441 if ((bound < std::numeric_limits<double>::max()) && (latency_ <= 0.0)) {
443 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), bound);
445 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), std::min(rate_, bound));
449 } // namespace simgrid::kernel::resource