1 /* Copyright (c) 2007-2023. 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/asserts.hpp>
9 #include <xbt/config.hpp>
11 #include "simgrid/config.h"
12 #include "src/kernel/EngineImpl.hpp"
13 #include "src/simgrid/math_utils.h"
14 #include "src/simgrid/module.hpp"
15 #if SIMGRID_HAVE_EIGEN3
16 #include "src/kernel/lmm/bmf.hpp"
18 #include "src/kernel/resource/models/ptask_L07.hpp"
19 #include "src/kernel/resource/profile/Event.hpp"
21 #include <unordered_set>
23 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(res_host);
24 XBT_LOG_EXTERNAL_CATEGORY(xbt_cfg);
29 static simgrid::config::Flag<std::string> cfg_ptask_solver("host/solver",
30 "Set linear equations solver used by ptask model",
32 &simgrid::kernel::lmm::System::validate_solver);
34 /**************************************/
35 /*** Resource Creation & Destruction **/
36 /**************************************/
37 SIMGRID_REGISTER_HOST_MODEL(
38 ptask_L07, "Host model somehow similar to Cas01+CM02+S19 but allowing parallel tasks", []() {
39 XBT_CINFO(xbt_cfg, "Switching to the L07 model to handle parallel tasks.");
40 xbt_assert(cfg_ptask_solver != "maxmin", "Invalid configuration. Cannot use maxmin solver with parallel tasks.");
42 xbt_assert(simgrid::config::is_default("network/model") && simgrid::config::is_default("cpu/model"),
43 "Changing the network or CPU model is not allowed when using the ptasks host model.");
45 auto* system = simgrid::kernel::lmm::System::build(cfg_ptask_solver.get(), true /* selective update */);
46 auto host_model = std::make_shared<simgrid::kernel::resource::HostL07Model>("Host_Ptask", system);
47 auto* engine = simgrid::kernel::EngineImpl::get_instance();
48 engine->add_model(host_model);
49 engine->get_netzone_root()->set_host_model(host_model);
52 namespace simgrid::kernel::resource {
54 HostL07Model::HostL07Model(const std::string& name, lmm::System* sys) : HostModel(name)
56 set_maxmin_system(sys);
58 auto net_model = std::make_shared<NetworkL07Model>("Network_Ptask", this, sys);
59 auto engine = EngineImpl::get_instance();
60 engine->add_model(net_model);
61 engine->get_netzone_root()->set_network_model(net_model);
63 auto cpu_model = std::make_shared<CpuL07Model>("Cpu_Ptask", this, sys);
64 engine->add_model(cpu_model);
65 engine->get_netzone_root()->set_cpu_pm_model(cpu_model);
67 simgrid_disk_models().by_name("S19").init();
70 CpuL07Model::CpuL07Model(const std::string& name, HostL07Model* hmodel, lmm::System* sys)
71 : CpuModel(name), hostModel_(hmodel)
73 set_maxmin_system(sys);
76 CpuL07Model::~CpuL07Model()
78 set_maxmin_system(nullptr);
81 NetworkL07Model::NetworkL07Model(const std::string& name, HostL07Model* hmodel, lmm::System* sys)
82 : NetworkModel(name), hostModel_(hmodel)
84 set_maxmin_system(sys);
85 loopback_.reset(create_link("__loopback__", {simgrid::config::get_value<double>("network/loopback-bw")}));
86 loopback_->set_sharing_policy(s4u::Link::SharingPolicy::FATPIPE, {});
87 loopback_->set_latency(simgrid::config::get_value<double>("network/loopback-lat"));
88 loopback_->get_iface()->seal();
91 NetworkL07Model::~NetworkL07Model()
93 set_maxmin_system(nullptr);
96 double HostL07Model::next_occurring_event(double now)
98 double min = HostModel::next_occurring_event_full(now);
99 for (Action const& action : *get_started_action_set()) {
100 const auto& net_action = static_cast<const L07Action&>(action);
101 if (net_action.get_latency() > 0 && (min < 0 || net_action.get_latency() < min)) {
102 min = net_action.get_latency();
103 XBT_DEBUG("Updating min with %p (start %f): %f", &net_action, net_action.get_start_time(), min);
106 XBT_DEBUG("min value: %f", min);
111 void HostL07Model::update_actions_state(double /*now*/, double delta)
113 for (auto it = std::begin(*get_started_action_set()); it != std::end(*get_started_action_set());) {
114 auto& action = static_cast<L07Action&>(*it);
115 ++it; // increment iterator here since the following calls to action.finish() may invalidate it
116 if (action.get_latency() > 0) {
117 if (action.get_latency() > delta) {
118 action.update_latency(delta, sg_precision_timing);
120 action.set_latency(0.0);
122 if ((action.get_latency() <= 0.0) && (action.is_suspended() == 0)) {
123 action.update_bound();
124 get_maxmin_system()->update_variable_penalty(action.get_variable(), 1.0);
125 action.set_last_update();
128 XBT_DEBUG("Action (%p) : remains (%g) updated by %g.", &action, action.get_remains(), action.get_rate() * delta);
129 action.update_remains(action.get_rate() * delta);
130 action.update_max_duration(delta);
132 XBT_DEBUG("Action (%p) : remains (%g).", &action, action.get_remains());
134 /* In the next if cascade, the action can be finished either because:
135 * - The amount of remaining work reached 0
136 * - The max duration was reached
137 * If it's not done, it may have failed.
140 if (((action.get_remains() <= 0) && (action.get_variable()->get_penalty() > 0)) ||
141 ((action.get_max_duration() != NO_MAX_DURATION) && (action.get_max_duration() <= 0))) {
142 action.finish(Action::State::FINISHED);
146 /* Need to check that none of the model has failed */
148 const lmm::Constraint* cnst = action.get_variable()->get_constraint(i);
149 while (cnst != nullptr) {
151 if (not cnst->get_id()->is_on()) {
152 XBT_DEBUG("Action (%p) Failed!!", &action);
153 action.finish(Action::State::FAILED);
156 cnst = action.get_variable()->get_constraint(i);
161 CpuAction* HostL07Model::execute_parallel(const std::vector<s4u::Host*>& host_list, const double* flops_amount,
162 const double* bytes_amount, double rate)
164 return new L07Action(this, host_list, flops_amount, bytes_amount, rate);
167 L07Action::L07Action(Model* model, const std::vector<s4u::Host*>& host_list, const double* flops_amount,
168 const double* bytes_amount, double rate)
169 : CpuAction(model, 1.0, false)
170 , host_list_(host_list)
171 , computation_amount_(flops_amount)
172 , communication_amount_(bytes_amount)
176 const size_t host_nb = host_list_.size();
177 size_t used_host_nb = 0; /* Only the hosts with something to compute (>0 flops) are counted) */
178 double latency = 0.0;
179 this->set_last_update();
181 if (flops_amount != nullptr)
182 used_host_nb += std::count_if(flops_amount, flops_amount + host_nb, [](double x) { return x > 0.0; });
184 /* Compute the number of affected resources... */
185 if (bytes_amount != nullptr) {
186 std::unordered_set<const char*> affected_links;
188 for (size_t k = 0; k < host_nb * host_nb; k++) {
189 if (bytes_amount[k] <= 0)
193 std::vector<StandardLinkImpl*> route;
194 host_list_[k / host_nb]->route_to(host_list_[k % host_nb], route, &lat);
195 latency = std::max(latency, lat);
197 for (auto const& link : route)
198 affected_links.insert(link->get_cname());
201 link_nb = affected_links.size();
204 XBT_DEBUG("Creating a parallel task (%p) with %zu hosts and %zu unique links.", this, host_nb, link_nb);
207 set_variable(model->get_maxmin_system()->variable_new(this, 1.0, (rate > 0 ? rate : -1.0), host_nb + link_nb));
210 model->get_maxmin_system()->update_variable_penalty(get_variable(), 0.0);
212 /* Expand it for the CPUs even if there is nothing to compute, to make sure that it gets expended even if there is no
213 * communication either */
214 for (size_t i = 0; i < host_nb; i++) {
215 model->get_maxmin_system()->expand(host_list[i]->get_cpu()->get_constraint(), get_variable(),
216 (flops_amount == nullptr ? 0.0 : flops_amount[i]), true);
219 if (bytes_amount != nullptr) {
220 for (size_t k = 0; k < host_nb * host_nb; k++) {
221 if (bytes_amount[k] <= 0.0)
223 std::vector<StandardLinkImpl*> route;
224 host_list_[k / host_nb]->route_to(host_list_[k % host_nb], route, nullptr);
226 for (auto const& link : route)
227 model->get_maxmin_system()->expand(link->get_constraint(), this->get_variable(), bytes_amount[k]);
231 if (link_nb + used_host_nb == 0) {
233 this->set_remains(0.0);
235 /* finally calculate the initial bound value */
239 Action* NetworkL07Model::communicate(s4u::Host* src, s4u::Host* dst, double size, double rate, bool /* streamed */)
241 std::vector<s4u::Host*> host_list = {src, dst};
242 const auto* flops_amount = new double[2]();
243 auto* bytes_amount = new double[4]();
245 bytes_amount[1] = size;
247 Action* res = hostModel_->execute_parallel(host_list, flops_amount, bytes_amount, rate);
248 static_cast<L07Action*>(res)->free_arrays_ = true;
252 CpuImpl* CpuL07Model::create_cpu(s4u::Host* host, const std::vector<double>& speed_per_pstate)
254 return (new CpuL07(host, speed_per_pstate))->set_model(this);
257 StandardLinkImpl* NetworkL07Model::create_link(const std::string& name, const std::vector<double>& bandwidths)
259 xbt_assert(bandwidths.size() == 1, "Non WIFI link must have only 1 bandwidth.");
260 auto link = new LinkL07(name, bandwidths[0], get_maxmin_system());
261 link->set_model(this);
265 StandardLinkImpl* NetworkL07Model::create_wifi_link(const std::string& name, const std::vector<double>& bandwidths)
274 CpuAction* CpuL07::execution_start(double size, double user_bound)
276 std::vector<s4u::Host*> host_list = {get_iface()};
277 xbt_assert(user_bound <= 0, "User bound not supported by ptask model");
279 auto* flops_amount = new double[host_list.size()]();
280 flops_amount[0] = size;
283 static_cast<CpuL07Model*>(get_model())->hostModel_->execute_parallel(host_list, flops_amount, nullptr, -1);
284 static_cast<L07Action*>(res)->free_arrays_ = true;
288 CpuAction* CpuL07::sleep(double duration)
290 auto* action = static_cast<L07Action*>(execution_start(1.0, -1));
291 action->set_max_duration(duration);
292 action->set_suspend_state(Action::SuspendStates::SLEEPING);
293 get_model()->get_maxmin_system()->update_variable_penalty(action->get_variable(), 0.0);
298 /** @brief take into account changes of speed (either load or max) */
299 void CpuL07::on_speed_change()
301 const lmm::Element* elem = nullptr;
303 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(),
304 get_core_count() * speed_.peak * speed_.scale);
306 while (const auto* var = get_constraint()->get_variable(&elem)) {
307 const auto* action = static_cast<L07Action*>(var->get_id());
308 action->update_bound();
311 CpuImpl::on_speed_change();
314 LinkL07::LinkL07(const std::string& name, double bandwidth, lmm::System* system) : StandardLinkImpl(name)
316 this->set_constraint(system->constraint_new(this, bandwidth));
317 bandwidth_.peak = bandwidth;
320 void CpuL07::apply_event(profile::Event* triggered, double value)
322 XBT_DEBUG("Updating cpu %s (%p) with value %g", get_cname(), this, value);
323 if (triggered == speed_.event) {
324 speed_.scale = value;
326 tmgr_trace_event_unref(&speed_.event);
328 } else if (triggered == get_state_event()) {
331 XBT_VERB("Restart actors on host %s", get_iface()->get_cname());
332 get_iface()->turn_on();
335 get_iface()->turn_off();
339 xbt_die("Unknown event!\n");
343 void LinkL07::apply_event(profile::Event* triggered, double value)
345 XBT_DEBUG("Updating link %s (%p) with value=%f", get_cname(), this, value);
346 if (triggered == bandwidth_.event) {
347 set_bandwidth(value);
348 tmgr_trace_event_unref(&bandwidth_.event);
350 } else if (triggered == latency_.event) {
352 tmgr_trace_event_unref(&latency_.event);
354 } else if (triggered == get_state_event()) {
361 xbt_die("Unknown event ! \n");
365 void LinkL07::set_bandwidth(double value)
367 bandwidth_.peak = value;
368 StandardLinkImpl::on_bandwidth_change();
370 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), bandwidth_.peak * bandwidth_.scale);
373 void LinkL07::set_latency(double value)
375 latency_check(value);
376 const lmm::Element* elem = nullptr;
378 latency_.peak = value;
379 while (const auto* var = get_constraint()->get_variable(&elem)) {
380 const auto* action = static_cast<L07Action*>(var->get_id());
381 action->update_bound();
384 LinkL07::~LinkL07() = default;
390 L07Action::~L07Action()
393 delete[] computation_amount_;
394 delete[] communication_amount_;
398 double L07Action::calculate_network_bound() const
400 double lat_current = 0.0;
401 double lat_bound = std::numeric_limits<double>::max();
403 size_t host_count = host_list_.size();
405 if (communication_amount_ == nullptr) {
409 for (size_t i = 0; i < host_count; i++) {
410 for (size_t j = 0; j < host_count; j++) {
411 if (communication_amount_[i * host_count + j] > 0) {
413 std::vector<StandardLinkImpl*> route;
414 host_list_.at(i)->route_to(host_list_.at(j), route, &lat);
416 lat_current = std::max(lat_current, lat * communication_amount_[i * host_count + j]);
420 if (lat_current > 0) {
421 lat_bound = NetworkModel::cfg_tcp_gamma / (2.0 * lat_current);
426 double L07Action::calculate_cpu_bound() const
428 double cpu_bound = std::numeric_limits<double>::max();
430 if (computation_amount_ == nullptr) {
434 for (size_t i = 0; i < host_list_.size(); i++) {
435 if (computation_amount_[i] > 0) {
436 cpu_bound = std::min(cpu_bound, host_list_[i]->get_cpu()->get_speed(1.0) *
437 host_list_[i]->get_cpu()->get_speed_ratio() / computation_amount_[i]);
443 void L07Action::update_bound() const
445 double bound = std::min(calculate_network_bound(), calculate_cpu_bound());
447 XBT_DEBUG("action (%p) : bound = %g", this, bound);
449 /* latency has been paid (or no latency), we can set the appropriate bound for multicore or network limit */
450 if ((bound < std::numeric_limits<double>::max()) && (latency_ <= 0.0)) {
452 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), bound);
454 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), std::min(rate_, bound));
458 } // namespace simgrid::kernel::resource