1 /* Copyright (c) 2007-2020. 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 "src/kernel/resource/profile/Event.hpp"
8 #include "surf/surf.hpp"
9 #include "xbt/config.hpp"
11 #include <unordered_set>
13 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(surf_host);
14 XBT_LOG_EXTERNAL_CATEGORY(xbt_cfg);
16 /**************************************/
17 /*** Resource Creation & Destruction **/
18 /**************************************/
19 void surf_host_model_init_ptask_L07()
21 XBT_CINFO(xbt_cfg,"Switching to the L07 model to handle parallel tasks.");
22 xbt_assert(not surf_cpu_model_pm, "Cannot switch to ptasks: CPU model already defined");
23 xbt_assert(not surf_network_model, "Cannot switch to ptasks: network model already defined");
25 surf_host_model = new simgrid::surf::HostL07Model();
26 all_existing_models.push_back(surf_host_model);
32 HostL07Model::HostL07Model() : HostModel() {
33 auto* maxmin_system = new simgrid::kernel::lmm::FairBottleneck(true /* selective update */);
34 set_maxmin_system(maxmin_system);
35 surf_network_model = new NetworkL07Model(this, maxmin_system);
36 surf_cpu_model_pm = new CpuL07Model(this, maxmin_system);
39 HostL07Model::~HostL07Model()
41 delete surf_network_model;
42 delete surf_cpu_model_pm;
45 CpuL07Model::CpuL07Model(HostL07Model* hmodel, kernel::lmm::System* sys)
46 : CpuModel(Model::UpdateAlgo::FULL), hostModel_(hmodel)
48 set_maxmin_system(sys);
51 CpuL07Model::~CpuL07Model()
53 set_maxmin_system(nullptr);
56 NetworkL07Model::NetworkL07Model(HostL07Model* hmodel, kernel::lmm::System* sys)
57 : NetworkModel(Model::UpdateAlgo::FULL), hostModel_(hmodel)
59 set_maxmin_system(sys);
60 loopback_ = NetworkL07Model::create_link("__loopback__",
61 std::vector<double>{simgrid::config::get_value<double>("network/loopback-bw")},
62 simgrid::config::get_value<double>("network/loopback-lat"),
63 s4u::Link::SharingPolicy::FATPIPE);
66 NetworkL07Model::~NetworkL07Model()
68 set_maxmin_system(nullptr);
71 double HostL07Model::next_occurring_event(double now)
73 double min = HostModel::next_occurring_event_full(now);
74 for (kernel::resource::Action const& action : *get_started_action_set()) {
75 const L07Action& net_action = static_cast<const L07Action&>(action);
76 if (net_action.get_latency() > 0 && (min < 0 || net_action.get_latency() < min)) {
77 min = net_action.get_latency();
78 XBT_DEBUG("Updating min with %p (start %f): %f", &net_action, net_action.get_start_time(), min);
81 XBT_DEBUG("min value: %f", min);
86 void HostL07Model::update_actions_state(double /*now*/, double delta)
88 for (auto it = std::begin(*get_started_action_set()); it != std::end(*get_started_action_set());) {
89 L07Action& action = static_cast<L07Action&>(*it);
90 ++it; // increment iterator here since the following calls to action.finish() may invalidate it
91 if (action.get_latency() > 0) {
92 if (action.get_latency() > delta) {
93 action.update_latency(delta, sg_surf_precision);
95 action.set_latency(0.0);
97 if ((action.get_latency() <= 0.0) && (action.is_suspended() == 0)) {
99 get_maxmin_system()->update_variable_penalty(action.get_variable(), 1.0);
100 action.set_last_update();
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);
106 action.update_max_duration(delta);
108 XBT_DEBUG("Action (%p) : remains (%g).", &action, action.get_remains());
110 /* In the next if cascade, the action can be finished either because:
111 * - The amount of remaining work reached 0
112 * - The max duration was reached
113 * If it's not done, it may have failed.
116 if (((action.get_remains() <= 0) && (action.get_variable()->get_penalty() > 0)) ||
117 ((action.get_max_duration() != NO_MAX_DURATION) && (action.get_max_duration() <= 0))) {
118 action.finish(kernel::resource::Action::State::FINISHED);
122 /* Need to check that none of the model has failed */
124 const kernel::lmm::Constraint* cnst = action.get_variable()->get_constraint(i);
125 while (cnst != nullptr) {
127 const kernel::resource::Resource* constraint_id = cnst->get_id();
128 if (not constraint_id->is_on()) {
129 XBT_DEBUG("Action (%p) Failed!!", &action);
130 action.finish(kernel::resource::Action::State::FAILED);
133 cnst = action.get_variable()->get_constraint(i);
138 kernel::resource::CpuAction* HostL07Model::execute_parallel(const std::vector<s4u::Host*>& host_list,
139 const double* flops_amount, const double* bytes_amount,
142 return new L07Action(this, host_list, flops_amount, bytes_amount, rate);
145 L07Action::L07Action(kernel::resource::Model* model, const std::vector<s4u::Host*>& host_list,
146 const double* flops_amount, const double* bytes_amount, double rate)
147 : CpuAction(model, 1.0, false), computationAmount_(flops_amount), communicationAmount_(bytes_amount), rate_(rate)
150 size_t used_host_nb = 0; /* Only the hosts with something to compute (>0 flops) are counted) */
151 double latency = 0.0;
152 this->set_last_update();
154 hostList_.insert(hostList_.end(), host_list.begin(), host_list.end());
156 if (flops_amount != nullptr)
157 used_host_nb += std::count_if(flops_amount, flops_amount + host_list.size(), [](double x) { return x > 0.0; });
159 /* Compute the number of affected resources... */
160 if(bytes_amount != nullptr) {
161 std::unordered_set<const char*> affected_links;
163 for (size_t k = 0; k < host_list.size() * host_list.size(); k++) {
164 if (bytes_amount[k] <= 0)
168 std::vector<kernel::resource::LinkImpl*> route;
169 hostList_[k / host_list.size()]->route_to(hostList_[k % host_list.size()], route, &lat);
170 latency = std::max(latency, lat);
172 for (auto const& link : route)
173 affected_links.insert(link->get_cname());
176 link_nb = affected_links.size();
179 XBT_DEBUG("Creating a parallel task (%p) with %zu hosts and %zu unique links.", this, host_list.size(), link_nb);
183 model->get_maxmin_system()->variable_new(this, 1.0, (rate > 0 ? rate : -1.0), host_list.size() + link_nb));
186 model->get_maxmin_system()->update_variable_penalty(get_variable(), 0.0);
188 /* Expand it for the CPUs even if there is nothing to compute, to make sure that it gets expended even if there is no
189 * communication either */
190 for (size_t i = 0; i < host_list.size(); i++)
191 model->get_maxmin_system()->expand(host_list[i]->pimpl_cpu->get_constraint(), get_variable(),
192 (flops_amount == nullptr ? 0.0 : flops_amount[i]));
194 if (bytes_amount != nullptr) {
195 for (size_t k = 0; k < host_list.size() * host_list.size(); k++) {
196 if (bytes_amount[k] <= 0.0)
198 std::vector<kernel::resource::LinkImpl*> route;
199 hostList_[k / host_list.size()]->route_to(hostList_[k % host_list.size()], route, nullptr);
201 for (auto const& link : route)
202 model->get_maxmin_system()->expand_add(link->get_constraint(), this->get_variable(), bytes_amount[k]);
206 if (link_nb + used_host_nb == 0) {
208 this->set_remains(0.0);
212 kernel::resource::Action* NetworkL07Model::communicate(s4u::Host* src, s4u::Host* dst, double size, double rate)
214 std::vector<s4u::Host*> host_list = {src, dst};
215 const double* flops_amount = new double[2]();
216 double* bytes_amount = new double[4]();
218 bytes_amount[1] = size;
220 kernel::resource::Action* res = hostModel_->execute_parallel(host_list, flops_amount, bytes_amount, rate);
221 static_cast<L07Action*>(res)->free_arrays_ = true;
225 kernel::resource::Cpu* CpuL07Model::create_cpu(s4u::Host* host, const std::vector<double>& speed_per_pstate, int core)
227 return new CpuL07(this, host, speed_per_pstate, core);
230 kernel::resource::LinkImpl* NetworkL07Model::create_link(const std::string& name, const std::vector<double>& bandwidths,
231 double latency, s4u::Link::SharingPolicy policy)
233 xbt_assert(bandwidths.size() == 1, "Non WIFI link must have only 1 bandwidth.");
234 return new LinkL07(this, name, bandwidths[0], latency, policy);
241 CpuL07::CpuL07(CpuL07Model* model, simgrid::s4u::Host* host, const 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(*get_iface());
262 kernel::resource::CpuAction* CpuL07::execution_start(double size)
264 std::vector<s4u::Host*> host_list = {get_host()};
266 double* flops_amount = new double[host_list.size()]();
267 flops_amount[0] = size;
269 kernel::resource::CpuAction* res =
270 static_cast<CpuL07Model*>(get_model())->hostModel_->execute_parallel(host_list, flops_amount, nullptr, -1);
271 static_cast<L07Action*>(res)->free_arrays_ = true;
275 kernel::resource::CpuAction* CpuL07::sleep(double duration)
277 L07Action *action = static_cast<L07Action*>(execution_start(1.0));
278 action->set_max_duration(duration);
279 action->set_suspend_state(kernel::resource::Action::SuspendStates::SLEEPING);
280 get_model()->get_maxmin_system()->update_variable_penalty(action->get_variable(), 0.0);
285 bool CpuL07::is_used()
287 return get_model()->get_maxmin_system()->constraint_used(get_constraint());
290 /** @brief take into account changes of speed (either load or max) */
291 void CpuL07::on_speed_change()
293 const kernel::lmm::Variable* var;
294 const kernel::lmm::Element* elem = nullptr;
296 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), speed_.peak * speed_.scale);
297 while ((var = get_constraint()->get_variable(&elem))) {
298 const kernel::resource::Action* action = var->get_id();
300 get_model()->get_maxmin_system()->update_variable_bound(action->get_variable(), speed_.scale * speed_.peak);
303 Cpu::on_speed_change();
306 bool LinkL07::is_used()
308 return get_model()->get_maxmin_system()->constraint_used(get_constraint());
311 void CpuL07::apply_event(kernel::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 == state_event_) {
322 XBT_VERB("Restart actors on host %s", get_host()->get_cname());
323 get_host()->turn_on();
326 get_host()->turn_off();
327 tmgr_trace_event_unref(&state_event_);
330 xbt_die("Unknown event!\n");
334 void LinkL07::apply_event(kernel::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 == state_event_) {
350 tmgr_trace_event_unref(&state_event_);
353 xbt_die("Unknown event ! \n");
357 void LinkL07::set_bandwidth(double value)
359 bandwidth_.peak = value;
360 LinkImpl::on_bandwidth_change();
362 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), bandwidth_.peak * bandwidth_.scale);
365 void LinkL07::set_latency(double value)
367 const kernel::lmm::Variable* var;
369 const kernel::lmm::Element* elem = nullptr;
371 latency_.peak = value;
372 while ((var = get_constraint()->get_variable(&elem))) {
373 action = static_cast<L07Action*>(var->get_id());
374 action->updateBound();
377 LinkL07::~LinkL07() = default;
383 L07Action::~L07Action()
386 delete[] computationAmount_;
387 delete[] communicationAmount_;
391 void L07Action::updateBound()
393 double lat_current = 0.0;
395 size_t host_count = hostList_.size();
397 if (communicationAmount_ != nullptr) {
398 for (size_t i = 0; i < host_count; i++) {
399 for (size_t j = 0; j < host_count; j++) {
400 if (communicationAmount_[i * host_count + j] > 0) {
402 std::vector<kernel::resource::LinkImpl*> route;
403 hostList_.at(i)->route_to(hostList_.at(j), route, &lat);
405 lat_current = std::max(lat_current, lat * communicationAmount_[i * host_count + j]);
410 double lat_bound = kernel::resource::NetworkModel::cfg_tcp_gamma / (2.0 * lat_current);
411 XBT_DEBUG("action (%p) : lat_bound = %g", this, lat_bound);
412 if ((latency_ <= 0.0) && is_running()) {
414 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), lat_bound);
416 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), std::min(rate_, lat_bound));