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. */
9 #include <unordered_set>
11 #include "ptask_L07.hpp"
13 #include "cpu_interface.hpp"
14 #include "xbt/utility.hpp"
16 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(surf_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.");
25 xbt_assert(not surf_cpu_model_pm, "CPU model type already defined");
26 xbt_assert(not surf_network_model, "network model type already defined");
28 surf_host_model = new simgrid::surf::HostL07Model();
29 all_existing_models->push_back(surf_host_model);
36 HostL07Model::HostL07Model() : HostModel() {
37 auto* maxmin_system = new simgrid::kernel::lmm::FairBottleneck(true /* selective update */);
38 set_maxmin_system(maxmin_system);
39 surf_network_model = new NetworkL07Model(this, maxmin_system);
40 surf_cpu_model_pm = new CpuL07Model(this, maxmin_system);
43 HostL07Model::~HostL07Model()
45 delete surf_network_model;
46 delete surf_cpu_model_pm;
49 CpuL07Model::CpuL07Model(HostL07Model* hmodel, kernel::lmm::System* sys)
50 : CpuModel(Model::UpdateAlgo::Full), hostModel_(hmodel)
52 set_maxmin_system(sys);
55 CpuL07Model::~CpuL07Model()
57 set_maxmin_system(nullptr);
60 NetworkL07Model::NetworkL07Model(HostL07Model* hmodel, kernel::lmm::System* sys)
61 : NetworkModel(Model::UpdateAlgo::Full), hostModel_(hmodel)
63 set_maxmin_system(sys);
64 loopback_ = NetworkL07Model::createLink("__loopback__", 498000000, 0.000015, SURF_LINK_FATPIPE);
67 NetworkL07Model::~NetworkL07Model()
69 set_maxmin_system(nullptr);
72 double HostL07Model::next_occuring_event(double now)
74 double min = HostModel::next_occuring_event_full(now);
75 for (kernel::resource::Action const& action : *get_running_action_set()) {
76 const L07Action& net_action = static_cast<const L07Action&>(action);
77 if (net_action.latency_ > 0 && (min < 0 || net_action.latency_ < min)) {
78 min = net_action.latency_;
79 XBT_DEBUG("Updating min with %p (start %f): %f", &net_action, net_action.get_start_time(), min);
82 XBT_DEBUG("min value: %f", min);
87 void HostL07Model::update_actions_state(double /*now*/, double delta)
89 for (auto it = std::begin(*get_running_action_set()); it != std::end(*get_running_action_set());) {
90 L07Action& action = static_cast<L07Action&>(*it);
91 ++it; // increment iterator here since the following calls to action.finish() may invalidate it
92 if (action.latency_ > 0) {
93 if (action.latency_ > delta) {
94 double_update(&(action.latency_), delta, sg_surf_precision);
96 action.latency_ = 0.0;
98 if ((action.latency_ <= 0.0) && (action.is_suspended() == 0)) {
100 get_maxmin_system()->update_variable_weight(action.get_variable(), 1.0);
103 XBT_DEBUG("Action (%p) : remains (%g) updated by %g.", &action, action.get_remains(),
104 action.get_variable()->get_value() * delta);
105 action.update_remains(action.get_variable()->get_value() * delta);
107 if (action.get_max_duration() > NO_MAX_DURATION)
108 action.update_max_duration(delta);
110 XBT_DEBUG("Action (%p) : remains (%g).", &action, action.get_remains());
112 /* In the next if cascade, the action can be finished either because:
113 * - The amount of remaining work reached 0
114 * - The max duration was reached
115 * If it's not done, it may have failed.
118 if (((action.get_remains() <= 0) && (action.get_variable()->get_weight() > 0)) ||
119 ((action.get_max_duration() > NO_MAX_DURATION) && (action.get_max_duration() <= 0))) {
120 action.finish(kernel::resource::Action::State::done);
122 /* Need to check that none of the model has failed */
124 kernel::lmm::Constraint* cnst = action.get_variable()->get_constraint(i);
125 while (cnst != nullptr) {
127 void* constraint_id = cnst->get_id();
128 if (static_cast<simgrid::kernel::resource::Resource*>(constraint_id)->isOff()) {
129 XBT_DEBUG("Action (%p) Failed!!", &action);
130 action.finish(kernel::resource::Action::State::failed);
133 cnst = action.get_variable()->get_constraint(i);
139 kernel::resource::Action* HostL07Model::execute_parallel(int host_nb, sg_host_t* host_list, double* flops_amount,
140 double* bytes_amount, double rate)
142 return new L07Action(this, host_nb, host_list, flops_amount, bytes_amount, rate);
145 L07Action::L07Action(kernel::resource::Model* model, int host_nb, sg_host_t* host_list, double* flops_amount,
146 double* bytes_amount, double rate)
147 : CpuAction(model, 1, 0), computationAmount_(flops_amount), communicationAmount_(bytes_amount), rate_(rate)
150 int nb_used_host = 0; /* Only the hosts with something to compute (>0 flops) are counted) */
151 double latency = 0.0;
153 this->hostList_->reserve(host_nb);
154 for (int i = 0; i < host_nb; i++) {
155 this->hostList_->push_back(host_list[i]);
156 if (flops_amount[i] > 0)
160 /* Compute the number of affected resources... */
161 if(bytes_amount != nullptr) {
162 std::unordered_set<const char*> affected_links;
164 for (int i = 0; i < host_nb; i++) {
165 for (int j = 0; j < host_nb; j++) {
167 if (bytes_amount[i * host_nb + j] > 0) {
170 std::vector<LinkImpl*> route;
171 hostList_->at(i)->routeTo(hostList_->at(j), route, &lat);
172 latency = std::max(latency, lat);
174 for (auto const& link : route)
175 affected_links.insert(link->get_cname());
180 nb_link = affected_links.size();
183 XBT_DEBUG("Creating a parallel task (%p) with %d hosts and %d unique links.", this, host_nb, nb_link);
186 set_variable(model->get_maxmin_system()->variable_new(this, 1.0, (rate > 0 ? rate : -1.0), host_nb + nb_link));
189 model->get_maxmin_system()->update_variable_weight(get_variable(), 0.0);
191 for (int i = 0; i < host_nb; i++)
192 model->get_maxmin_system()->expand(host_list[i]->pimpl_cpu->constraint(), get_variable(), flops_amount[i]);
194 if(bytes_amount != nullptr) {
195 for (int i = 0; i < host_nb; i++) {
196 for (int j = 0; j < host_nb; j++) {
197 if (bytes_amount[i * host_nb + j] > 0.0) {
198 std::vector<LinkImpl*> route;
199 hostList_->at(i)->routeTo(hostList_->at(j), route, nullptr);
201 for (auto const& link : route)
202 model->get_maxmin_system()->expand_add(link->constraint(), this->get_variable(),
203 bytes_amount[i * host_nb + j]);
209 if (nb_link + nb_used_host == 0) {
211 this->set_remains(0.0);
216 kernel::resource::Action* NetworkL07Model::communicate(s4u::Host* src, s4u::Host* dst, double size, double rate)
218 sg_host_t* host_list = new sg_host_t[2]();
219 double* flops_amount = new double[2]();
220 double* bytes_amount = new double[4]();
224 bytes_amount[1] = size;
226 return hostModel_->execute_parallel(2, host_list, flops_amount, bytes_amount, rate);
229 Cpu *CpuL07Model::createCpu(simgrid::s4u::Host *host, std::vector<double> *speedPerPstate, int core)
231 return new CpuL07(this, host, speedPerPstate, core);
234 LinkImpl* NetworkL07Model::createLink(const std::string& name, double bandwidth, double latency,
235 e_surf_link_sharing_policy_t policy)
237 return new LinkL07(this, name, bandwidth, latency, policy);
244 CpuL07::CpuL07(CpuL07Model* model, simgrid::s4u::Host* host, std::vector<double>* speedPerPstate, int core)
245 : Cpu(model, host, model->get_maxmin_system()->constraint_new(this, speedPerPstate->front()), speedPerPstate, core)
249 CpuL07::~CpuL07()=default;
251 LinkL07::LinkL07(NetworkL07Model* model, const std::string& name, double bandwidth, double latency,
252 e_surf_link_sharing_policy_t policy)
253 : LinkImpl(model, name, model->get_maxmin_system()->constraint_new(this, bandwidth))
255 bandwidth_.peak = bandwidth;
256 latency_.peak = latency;
258 if (policy == SURF_LINK_FATPIPE)
259 constraint()->unshare();
261 s4u::Link::onCreation(this->piface_);
264 kernel::resource::Action* CpuL07::execution_start(double size)
266 sg_host_t* host_list = new sg_host_t[1]();
267 double* flops_amount = new double[1]();
269 host_list[0] = getHost();
270 flops_amount[0] = size;
272 return static_cast<CpuL07Model*>(model())->hostModel_->execute_parallel(1, host_list, flops_amount, nullptr, -1);
275 kernel::resource::Action* CpuL07::sleep(double duration)
277 L07Action *action = static_cast<L07Action*>(execution_start(1.0));
278 action->set_max_duration(duration);
279 action->suspended_ = kernel::resource::Action::SuspendStates::sleeping;
280 model()->get_maxmin_system()->update_variable_weight(action->get_variable(), 0.0);
285 bool CpuL07::is_used()
287 return model()->get_maxmin_system()->constraint_used(constraint());
290 /** @brief take into account changes of speed (either load or max) */
291 void CpuL07::onSpeedChange() {
292 kernel::lmm::Variable* var = nullptr;
293 const_lmm_element_t elem = nullptr;
295 model()->get_maxmin_system()->update_constraint_bound(constraint(), speed_.peak * speed_.scale);
296 while ((var = constraint()->get_variable(&elem))) {
297 kernel::resource::Action* action = static_cast<kernel::resource::Action*>(var->get_id());
299 model()->get_maxmin_system()->update_variable_bound(action->get_variable(), speed_.scale * speed_.peak);
302 Cpu::onSpeedChange();
305 bool LinkL07::is_used()
307 return model()->get_maxmin_system()->constraint_used(constraint());
310 void CpuL07::apply_event(tmgr_trace_event_t triggered, double value)
312 XBT_DEBUG("Updating cpu %s (%p) with value %g", get_cname(), this, value);
313 if (triggered == speed_.event) {
314 speed_.scale = value;
316 tmgr_trace_event_unref(&speed_.event);
318 } else if (triggered == stateEvent_) {
323 tmgr_trace_event_unref(&stateEvent_);
326 xbt_die("Unknown event!\n");
330 void LinkL07::apply_event(tmgr_trace_event_t triggered, double value)
332 XBT_DEBUG("Updating link %s (%p) with value=%f", get_cname(), this, value);
333 if (triggered == bandwidth_.event) {
335 tmgr_trace_event_unref(&bandwidth_.event);
337 } else if (triggered == latency_.event) {
339 tmgr_trace_event_unref(&latency_.event);
341 } else if (triggered == stateEvent_) {
346 tmgr_trace_event_unref(&stateEvent_);
349 xbt_die("Unknown event ! \n");
353 void LinkL07::setBandwidth(double value)
355 bandwidth_.peak = value;
356 model()->get_maxmin_system()->update_constraint_bound(constraint(), bandwidth_.peak * bandwidth_.scale);
359 void LinkL07::setLatency(double value)
361 kernel::lmm::Variable* var = nullptr;
363 const_lmm_element_t elem = nullptr;
365 latency_.peak = value;
366 while ((var = constraint()->get_variable(&elem))) {
367 action = static_cast<L07Action*>(var->get_id());
368 action->updateBound();
371 LinkL07::~LinkL07() = default;
377 L07Action::~L07Action(){
379 delete[] communicationAmount_;
380 delete[] computationAmount_;
383 void L07Action::updateBound()
385 double lat_current = 0.0;
387 int hostNb = hostList_->size();
389 if (communicationAmount_ != nullptr) {
390 for (int i = 0; i < hostNb; i++) {
391 for (int j = 0; j < hostNb; j++) {
393 if (communicationAmount_[i * hostNb + j] > 0) {
395 std::vector<LinkImpl*> route;
396 hostList_->at(i)->routeTo(hostList_->at(j), route, &lat);
398 lat_current = std::max(lat_current, lat * communicationAmount_[i * hostNb + j]);
403 double lat_bound = sg_tcp_gamma / (2.0 * lat_current);
404 XBT_DEBUG("action (%p) : lat_bound = %g", this, lat_bound);
405 if ((latency_ <= 0.0) && (suspended_ == Action::SuspendStates::not_suspended)) {
407 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), lat_bound);
409 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), std::min(rate_, lat_bound));