1 /* Copyright (c) 2007-2019. 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(size_t 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, size_t 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 size_t used_host_nb = 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 (size_t i = 0; i < host_nb; i++)
152 this->hostList_.push_back(host_list[i]);
154 if (flops_amount != nullptr)
155 for (size_t i = 0; i < host_nb; i++)
156 if (flops_amount[i] > 0)
159 /* Compute the number of affected resources... */
160 if(bytes_amount != nullptr) {
161 std::unordered_set<const char*> affected_links;
163 for (size_t i = 0; i < host_nb; i++) {
164 for (size_t j = 0; j < host_nb; j++) {
166 if (bytes_amount[i * host_nb + j] > 0) {
169 std::vector<kernel::resource::LinkImpl*> route;
170 hostList_.at(i)->route_to(hostList_.at(j), route, &lat);
171 latency = std::max(latency, lat);
173 for (auto const& link : route)
174 affected_links.insert(link->get_cname());
179 link_nb = affected_links.size();
182 XBT_DEBUG("Creating a parallel task (%p) with %zu hosts and %zu unique links.", this, host_nb, link_nb);
185 set_variable(model->get_maxmin_system()->variable_new(this, 1.0, (rate > 0 ? rate : -1.0), host_nb + link_nb));
188 model->get_maxmin_system()->update_variable_weight(get_variable(), 0.0);
190 /* Expend it for the CPUs even if there is nothing to compute, to make sure that it gets expended even if there is no
191 * communication either */
192 for (size_t i = 0; i < host_nb; i++)
193 model->get_maxmin_system()->expand(host_list[i]->pimpl_cpu->get_constraint(), get_variable(),
194 (flops_amount == nullptr ? 0.0 : flops_amount[i]));
196 if (bytes_amount != nullptr) {
197 for (size_t i = 0; i < host_nb; i++) {
198 for (size_t j = 0; j < host_nb; j++) {
199 if (bytes_amount[i * host_nb + j] > 0.0) {
200 std::vector<kernel::resource::LinkImpl*> route;
201 hostList_.at(i)->route_to(hostList_.at(j), route, nullptr);
203 for (auto const& link : route)
204 model->get_maxmin_system()->expand_add(link->get_constraint(), this->get_variable(),
205 bytes_amount[i * host_nb + j]);
211 if (link_nb + used_host_nb == 0) {
213 this->set_remains(0.0);
217 kernel::resource::Action* NetworkL07Model::communicate(s4u::Host* src, s4u::Host* dst, double size, double rate)
219 sg_host_t* host_list = new sg_host_t[2]();
220 double* flops_amount = new double[2]();
221 double* bytes_amount = new double[4]();
225 bytes_amount[1] = size;
227 kernel::resource::Action* res = hostModel_->execute_parallel(2, host_list, flops_amount, bytes_amount, rate);
228 static_cast<L07Action*>(res)->free_arrays_ = true;
232 Cpu* CpuL07Model::create_cpu(simgrid::s4u::Host* host, std::vector<double>* speed_per_pstate, int core)
234 return new CpuL07(this, host, speed_per_pstate, core);
237 kernel::resource::LinkImpl* NetworkL07Model::create_link(const std::string& name, double bandwidth, double latency,
238 s4u::Link::SharingPolicy policy)
240 return new LinkL07(this, name, bandwidth, latency, policy);
247 CpuL07::CpuL07(CpuL07Model* model, simgrid::s4u::Host* host, std::vector<double>* speed_per_pstate, int core)
248 : Cpu(model, host, model->get_maxmin_system()->constraint_new(this, speed_per_pstate->front()), speed_per_pstate,
253 CpuL07::~CpuL07()=default;
255 LinkL07::LinkL07(NetworkL07Model* model, const std::string& name, double bandwidth, double latency,
256 s4u::Link::SharingPolicy policy)
257 : LinkImpl(model, name, model->get_maxmin_system()->constraint_new(this, bandwidth))
259 bandwidth_.peak = bandwidth;
260 latency_.peak = latency;
262 if (policy == s4u::Link::SharingPolicy::FATPIPE)
263 get_constraint()->unshare();
265 s4u::Link::on_creation(this->piface_);
268 kernel::resource::Action* CpuL07::execution_start(double size)
270 sg_host_t host_list[1] = {get_host()};
272 double* flops_amount = new double[1]();
273 flops_amount[0] = size;
275 kernel::resource::Action* res =
276 static_cast<CpuL07Model*>(get_model())->hostModel_->execute_parallel(1, host_list, flops_amount, nullptr, -1);
277 static_cast<L07Action*>(res)->free_arrays_ = true;
281 kernel::resource::Action* CpuL07::sleep(double duration)
283 L07Action *action = static_cast<L07Action*>(execution_start(1.0));
284 action->set_max_duration(duration);
285 action->suspended_ = kernel::resource::Action::SuspendStates::sleeping;
286 get_model()->get_maxmin_system()->update_variable_weight(action->get_variable(), 0.0);
291 bool CpuL07::is_used()
293 return get_model()->get_maxmin_system()->constraint_used(get_constraint());
296 /** @brief take into account changes of speed (either load or max) */
297 void CpuL07::on_speed_change()
299 kernel::lmm::Variable* var = nullptr;
300 const kernel::lmm::Element* elem = nullptr;
302 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), speed_.peak * speed_.scale);
303 while ((var = get_constraint()->get_variable(&elem))) {
304 kernel::resource::Action* action = static_cast<kernel::resource::Action*>(var->get_id());
306 get_model()->get_maxmin_system()->update_variable_bound(action->get_variable(), speed_.scale * speed_.peak);
309 Cpu::on_speed_change();
312 bool LinkL07::is_used()
314 return get_model()->get_maxmin_system()->constraint_used(get_constraint());
317 void CpuL07::apply_event(tmgr_trace_event_t triggered, double value)
319 XBT_DEBUG("Updating cpu %s (%p) with value %g", get_cname(), this, value);
320 if (triggered == speed_.event) {
321 speed_.scale = value;
323 tmgr_trace_event_unref(&speed_.event);
325 } else if (triggered == state_event_) {
328 XBT_VERB("Restart processes on host %s", get_host()->get_cname());
329 get_host()->turn_on();
332 get_host()->turn_off();
333 tmgr_trace_event_unref(&state_event_);
336 xbt_die("Unknown event!\n");
340 void LinkL07::apply_event(tmgr_trace_event_t triggered, double value)
342 XBT_DEBUG("Updating link %s (%p) with value=%f", get_cname(), this, value);
343 if (triggered == bandwidth_.event) {
344 set_bandwidth(value);
345 tmgr_trace_event_unref(&bandwidth_.event);
347 } else if (triggered == latency_.event) {
349 tmgr_trace_event_unref(&latency_.event);
351 } else if (triggered == state_event_) {
356 tmgr_trace_event_unref(&state_event_);
359 xbt_die("Unknown event ! \n");
363 void LinkL07::set_bandwidth(double value)
365 bandwidth_.peak = value;
366 LinkImpl::on_bandwidth_change();
368 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), bandwidth_.peak * bandwidth_.scale);
371 void LinkL07::set_latency(double value)
373 kernel::lmm::Variable* var = nullptr;
375 const kernel::lmm::Element* elem = nullptr;
377 latency_.peak = value;
378 while ((var = get_constraint()->get_variable(&elem))) {
379 action = static_cast<L07Action*>(var->get_id());
380 action->updateBound();
383 LinkL07::~LinkL07() = default;
389 L07Action::~L07Action()
392 delete[] computationAmount_;
393 delete[] communicationAmount_;
397 void L07Action::updateBound()
399 double lat_current = 0.0;
401 size_t host_count = hostList_.size();
403 if (communicationAmount_ != nullptr) {
404 for (size_t i = 0; i < host_count; i++) {
405 for (size_t j = 0; j < host_count; j++) {
407 if (communicationAmount_[i * host_count + j] > 0) {
409 std::vector<kernel::resource::LinkImpl*> route;
410 hostList_.at(i)->route_to(hostList_.at(j), route, &lat);
412 lat_current = std::max(lat_current, lat * communicationAmount_[i * host_count + j]);
417 double lat_bound = kernel::resource::NetworkModel::cfg_tcp_gamma / (2.0 * lat_current);
418 XBT_DEBUG("action (%p) : lat_bound = %g", this, lat_bound);
419 if ((latency_ <= 0.0) && (suspended_ == Action::SuspendStates::not_suspended)) {
421 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), lat_bound);
423 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), std::min(rate_, lat_bound));