1 /* Copyright (c) 2007-2018. 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 "surf/surf.hpp"
8 #include "xbt/config.hpp"
10 #include <unordered_set>
12 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(surf_host);
13 XBT_LOG_EXTERNAL_CATEGORY(xbt_cfg);
15 /**************************************/
16 /*** Resource Creation & Destruction **/
17 /**************************************/
18 void surf_host_model_init_ptask_L07()
20 XBT_CINFO(xbt_cfg,"Switching to the L07 model to handle parallel tasks.");
21 xbt_assert(not surf_cpu_model_pm, "Cannot switch to ptasks: CPU model already defined");
22 xbt_assert(not surf_network_model, "Cannot switch to ptasks: network model already defined");
24 surf_host_model = new simgrid::surf::HostL07Model();
25 all_existing_models.push_back(surf_host_model);
31 HostL07Model::HostL07Model() : HostModel() {
32 auto* maxmin_system = new simgrid::kernel::lmm::FairBottleneck(true /* selective update */);
33 set_maxmin_system(maxmin_system);
34 surf_network_model = new NetworkL07Model(this, maxmin_system);
35 surf_cpu_model_pm = new CpuL07Model(this, maxmin_system);
38 HostL07Model::~HostL07Model()
40 delete surf_network_model;
41 delete surf_cpu_model_pm;
44 CpuL07Model::CpuL07Model(HostL07Model* hmodel, kernel::lmm::System* sys)
45 : CpuModel(Model::UpdateAlgo::FULL), hostModel_(hmodel)
47 set_maxmin_system(sys);
50 CpuL07Model::~CpuL07Model()
52 set_maxmin_system(nullptr);
55 NetworkL07Model::NetworkL07Model(HostL07Model* hmodel, kernel::lmm::System* sys)
56 : NetworkModel(Model::UpdateAlgo::FULL), hostModel_(hmodel)
58 set_maxmin_system(sys);
59 loopback_ = NetworkL07Model::create_link("__loopback__", 498000000, 0.000015, s4u::Link::SharingPolicy::FATPIPE);
62 NetworkL07Model::~NetworkL07Model()
64 set_maxmin_system(nullptr);
67 double HostL07Model::next_occuring_event(double now)
69 double min = HostModel::next_occuring_event_full(now);
70 for (kernel::resource::Action const& action : *get_started_action_set()) {
71 const L07Action& net_action = static_cast<const L07Action&>(action);
72 if (net_action.latency_ > 0 && (min < 0 || net_action.latency_ < min)) {
73 min = net_action.latency_;
74 XBT_DEBUG("Updating min with %p (start %f): %f", &net_action, net_action.get_start_time(), min);
77 XBT_DEBUG("min value: %f", min);
82 void HostL07Model::update_actions_state(double /*now*/, double delta)
84 for (auto it = std::begin(*get_started_action_set()); it != std::end(*get_started_action_set());) {
85 L07Action& action = static_cast<L07Action&>(*it);
86 ++it; // increment iterator here since the following calls to action.finish() may invalidate it
87 if (action.latency_ > 0) {
88 if (action.latency_ > delta) {
89 double_update(&(action.latency_), delta, sg_surf_precision);
91 action.latency_ = 0.0;
93 if ((action.latency_ <= 0.0) && (action.is_suspended() == 0)) {
95 get_maxmin_system()->update_variable_weight(action.get_variable(), 1.0);
96 action.set_last_update();
99 XBT_DEBUG("Action (%p) : remains (%g) updated by %g.", &action, action.get_remains(),
100 action.get_variable()->get_value() * delta);
101 action.update_remains(action.get_variable()->get_value() * delta);
103 if (action.get_max_duration() > NO_MAX_DURATION)
104 action.update_max_duration(delta);
106 XBT_DEBUG("Action (%p) : remains (%g).", &action, action.get_remains());
108 /* In the next if cascade, the action can be finished either because:
109 * - The amount of remaining work reached 0
110 * - The max duration was reached
111 * If it's not done, it may have failed.
114 if (((action.get_remains() <= 0) && (action.get_variable()->get_weight() > 0)) ||
115 ((action.get_max_duration() > NO_MAX_DURATION) && (action.get_max_duration() <= 0))) {
116 action.finish(kernel::resource::Action::State::FINISHED);
118 /* Need to check that none of the model has failed */
120 kernel::lmm::Constraint* cnst = action.get_variable()->get_constraint(i);
121 while (cnst != nullptr) {
123 void* constraint_id = cnst->get_id();
124 if (static_cast<simgrid::kernel::resource::Resource*>(constraint_id)->is_off()) {
125 XBT_DEBUG("Action (%p) Failed!!", &action);
126 action.finish(kernel::resource::Action::State::FAILED);
129 cnst = action.get_variable()->get_constraint(i);
135 kernel::resource::Action* HostL07Model::execute_parallel(int host_nb, sg_host_t* host_list, double* flops_amount,
136 double* bytes_amount, double rate)
138 return new L07Action(this, host_nb, host_list, flops_amount, bytes_amount, rate);
141 L07Action::L07Action(kernel::resource::Model* model, int host_nb, sg_host_t* host_list, double* flops_amount,
142 double* bytes_amount, double rate)
143 : CpuAction(model, 1, 0), computationAmount_(flops_amount), communicationAmount_(bytes_amount), rate_(rate)
146 int nb_used_host = 0; /* Only the hosts with something to compute (>0 flops) are counted) */
147 double latency = 0.0;
148 this->set_last_update();
150 this->hostList_->reserve(host_nb);
151 for (int i = 0; i < host_nb; i++) {
152 this->hostList_->push_back(host_list[i]);
153 if (flops_amount[i] > 0)
157 /* Compute the number of affected resources... */
158 if(bytes_amount != nullptr) {
159 std::unordered_set<const char*> affected_links;
161 for (int i = 0; i < host_nb; i++) {
162 for (int j = 0; j < host_nb; j++) {
164 if (bytes_amount[i * host_nb + j] > 0) {
167 std::vector<kernel::resource::LinkImpl*> route;
168 hostList_->at(i)->route_to(hostList_->at(j), route, &lat);
169 latency = std::max(latency, lat);
171 for (auto const& link : route)
172 affected_links.insert(link->get_cname());
177 nb_link = affected_links.size();
180 XBT_DEBUG("Creating a parallel task (%p) with %d hosts and %d unique links.", this, host_nb, nb_link);
183 set_variable(model->get_maxmin_system()->variable_new(this, 1.0, (rate > 0 ? rate : -1.0), host_nb + nb_link));
186 model->get_maxmin_system()->update_variable_weight(get_variable(), 0.0);
188 for (int i = 0; i < host_nb; i++)
189 model->get_maxmin_system()->expand(host_list[i]->pimpl_cpu->get_constraint(), get_variable(), flops_amount[i]);
191 if(bytes_amount != nullptr) {
192 for (int i = 0; i < host_nb; i++) {
193 for (int j = 0; j < host_nb; j++) {
194 if (bytes_amount[i * host_nb + j] > 0.0) {
195 std::vector<kernel::resource::LinkImpl*> route;
196 hostList_->at(i)->route_to(hostList_->at(j), route, nullptr);
198 for (auto const& link : route)
199 model->get_maxmin_system()->expand_add(link->get_constraint(), this->get_variable(),
200 bytes_amount[i * host_nb + j]);
206 if (nb_link + nb_used_host == 0) {
208 this->set_remains(0.0);
213 kernel::resource::Action* NetworkL07Model::communicate(s4u::Host* src, s4u::Host* dst, double size, double rate)
215 sg_host_t* host_list = new sg_host_t[2]();
216 double* flops_amount = new double[2]();
217 double* bytes_amount = new double[4]();
221 bytes_amount[1] = size;
223 return hostModel_->execute_parallel(2, host_list, flops_amount, bytes_amount, rate);
226 Cpu* CpuL07Model::create_cpu(simgrid::s4u::Host* host, std::vector<double>* speed_per_pstate, int core)
228 return new CpuL07(this, host, speed_per_pstate, core);
231 kernel::resource::LinkImpl* NetworkL07Model::create_link(const std::string& name, double bandwidth, double latency,
232 s4u::Link::SharingPolicy policy)
234 return new LinkL07(this, name, bandwidth, latency, policy);
241 CpuL07::CpuL07(CpuL07Model* model, simgrid::s4u::Host* host, std::vector<double>* speed_per_pstate, int core)
242 : Cpu(model, host, model->get_maxmin_system()->constraint_new(this, speed_per_pstate->front()), speed_per_pstate,
247 CpuL07::~CpuL07()=default;
249 LinkL07::LinkL07(NetworkL07Model* model, const std::string& name, double bandwidth, double latency,
250 s4u::Link::SharingPolicy policy)
251 : LinkImpl(model, name, model->get_maxmin_system()->constraint_new(this, bandwidth))
253 bandwidth_.peak = bandwidth;
254 latency_.peak = latency;
256 if (policy == s4u::Link::SharingPolicy::FATPIPE)
257 get_constraint()->unshare();
259 s4u::Link::on_creation(this->piface_);
262 kernel::resource::Action* CpuL07::execution_start(double size)
264 sg_host_t* host_list = new sg_host_t[1]();
265 double* flops_amount = new double[1]();
267 host_list[0] = get_host();
268 flops_amount[0] = size;
270 return static_cast<CpuL07Model*>(get_model())->hostModel_->execute_parallel(1, host_list, flops_amount, nullptr, -1);
273 kernel::resource::Action* CpuL07::sleep(double duration)
275 L07Action *action = static_cast<L07Action*>(execution_start(1.0));
276 action->set_max_duration(duration);
277 action->suspended_ = kernel::resource::Action::SuspendStates::sleeping;
278 get_model()->get_maxmin_system()->update_variable_weight(action->get_variable(), 0.0);
283 bool CpuL07::is_used()
285 return get_model()->get_maxmin_system()->constraint_used(get_constraint());
288 /** @brief take into account changes of speed (either load or max) */
289 void CpuL07::on_speed_change()
291 kernel::lmm::Variable* var = nullptr;
292 const kernel::lmm::Element* elem = nullptr;
294 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), speed_.peak * speed_.scale);
295 while ((var = get_constraint()->get_variable(&elem))) {
296 kernel::resource::Action* action = static_cast<kernel::resource::Action*>(var->get_id());
298 get_model()->get_maxmin_system()->update_variable_bound(action->get_variable(), speed_.scale * speed_.peak);
301 Cpu::on_speed_change();
304 bool LinkL07::is_used()
306 return get_model()->get_maxmin_system()->constraint_used(get_constraint());
309 void CpuL07::apply_event(tmgr_trace_event_t 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_) {
322 tmgr_trace_event_unref(&state_event_);
325 xbt_die("Unknown event!\n");
329 void LinkL07::apply_event(tmgr_trace_event_t triggered, double value)
331 XBT_DEBUG("Updating link %s (%p) with value=%f", get_cname(), this, value);
332 if (triggered == bandwidth_.event) {
333 set_bandwidth(value);
334 tmgr_trace_event_unref(&bandwidth_.event);
336 } else if (triggered == latency_.event) {
338 tmgr_trace_event_unref(&latency_.event);
340 } else if (triggered == state_event_) {
345 tmgr_trace_event_unref(&state_event_);
348 xbt_die("Unknown event ! \n");
352 void LinkL07::set_bandwidth(double value)
354 bandwidth_.peak = value;
355 LinkImpl::on_bandwidth_change();
357 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), bandwidth_.peak * bandwidth_.scale);
360 void LinkL07::set_latency(double value)
362 kernel::lmm::Variable* var = nullptr;
364 const kernel::lmm::Element* elem = nullptr;
366 latency_.peak = value;
367 while ((var = get_constraint()->get_variable(&elem))) {
368 action = static_cast<L07Action*>(var->get_id());
369 action->updateBound();
372 LinkL07::~LinkL07() = default;
378 L07Action::~L07Action(){
380 delete[] communicationAmount_;
381 delete[] computationAmount_;
384 void L07Action::updateBound()
386 double lat_current = 0.0;
388 int hostNb = hostList_->size();
390 if (communicationAmount_ != nullptr) {
391 for (int i = 0; i < hostNb; i++) {
392 for (int j = 0; j < hostNb; j++) {
394 if (communicationAmount_[i * hostNb + j] > 0) {
396 std::vector<kernel::resource::LinkImpl*> route;
397 hostList_->at(i)->route_to(hostList_->at(j), route, &lat);
399 lat_current = std::max(lat_current, lat * communicationAmount_[i * hostNb + j]);
404 double lat_bound = kernel::resource::NetworkModel::cfg_tcp_gamma / (2.0 * lat_current);
405 XBT_DEBUG("action (%p) : lat_bound = %g", this, lat_bound);
406 if ((latency_ <= 0.0) && (suspended_ == Action::SuspendStates::not_suspended)) {
408 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), lat_bound);
410 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), std::min(rate_, lat_bound));