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);
102 action.update_max_duration(delta);
104 XBT_DEBUG("Action (%p) : remains (%g).", &action, action.get_remains());
106 /* In the next if cascade, the action can be finished either because:
107 * - The amount of remaining work reached 0
108 * - The max duration was reached
109 * If it's not done, it may have failed.
112 if (((action.get_remains() <= 0) && (action.get_variable()->get_weight() > 0)) ||
113 ((action.get_max_duration() != NO_MAX_DURATION) && (action.get_max_duration() <= 0))) {
114 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 (not static_cast<simgrid::kernel::resource::Resource*>(constraint_id)->is_on()) {
125 XBT_DEBUG("Action (%p) Failed!!", &action);
126 action.finish(kernel::resource::Action::State::FAILED);
129 cnst = action.get_variable()->get_constraint(i);
134 kernel::resource::CpuAction* HostL07Model::execute_parallel(const std::vector<s4u::Host*>& host_list,
135 const double* flops_amount, const double* bytes_amount,
138 return new L07Action(this, host_list, flops_amount, bytes_amount, rate);
141 L07Action::L07Action(kernel::resource::Model* model, const std::vector<s4u::Host*>& host_list,
142 const double* flops_amount, const 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 hostList_.insert(hostList_.end(), host_list.begin(), host_list.end());
152 if (flops_amount != nullptr)
153 used_host_nb += std::count_if(flops_amount, flops_amount + host_list.size(), [](double x) { return x > 0.0; });
155 /* Compute the number of affected resources... */
156 if(bytes_amount != nullptr) {
157 std::unordered_set<const char*> affected_links;
159 for (size_t k = 0; k < host_list.size() * host_list.size(); k++) {
160 if (bytes_amount[k] <= 0)
164 std::vector<kernel::resource::LinkImpl*> route;
165 hostList_[k / host_list.size()]->route_to(hostList_[k % host_list.size()], route, &lat);
166 latency = std::max(latency, lat);
168 for (auto const& link : route)
169 affected_links.insert(link->get_cname());
172 link_nb = affected_links.size();
175 XBT_DEBUG("Creating a parallel task (%p) with %zu hosts and %zu unique links.", this, host_list.size(), link_nb);
179 model->get_maxmin_system()->variable_new(this, 1.0, (rate > 0 ? rate : -1.0), host_list.size() + link_nb));
182 model->get_maxmin_system()->update_variable_weight(get_variable(), 0.0);
184 /* Expand it for the CPUs even if there is nothing to compute, to make sure that it gets expended even if there is no
185 * communication either */
186 for (size_t i = 0; i < host_list.size(); i++)
187 model->get_maxmin_system()->expand(host_list[i]->pimpl_cpu->get_constraint(), get_variable(),
188 (flops_amount == nullptr ? 0.0 : flops_amount[i]));
190 if (bytes_amount != nullptr) {
191 for (size_t k = 0; k < host_list.size() * host_list.size(); k++) {
192 if (bytes_amount[k] <= 0.0)
194 std::vector<kernel::resource::LinkImpl*> route;
195 hostList_[k / host_list.size()]->route_to(hostList_[k % host_list.size()], route, nullptr);
197 for (auto const& link : route)
198 model->get_maxmin_system()->expand_add(link->get_constraint(), this->get_variable(), bytes_amount[k]);
202 if (link_nb + used_host_nb == 0) {
204 this->set_remains(0.0);
208 kernel::resource::Action* NetworkL07Model::communicate(s4u::Host* src, s4u::Host* dst, double size, double rate)
210 std::vector<s4u::Host*> host_list = {src, dst};
211 double* flops_amount = new double[2]();
212 double* bytes_amount = new double[4]();
214 bytes_amount[1] = size;
216 kernel::resource::Action* res = hostModel_->execute_parallel(host_list, flops_amount, bytes_amount, rate);
217 static_cast<L07Action*>(res)->free_arrays_ = true;
221 kernel::resource::Cpu* CpuL07Model::create_cpu(s4u::Host* host, const std::vector<double>& speed_per_pstate, int core)
223 return new CpuL07(this, host, speed_per_pstate, core);
226 kernel::resource::LinkImpl* NetworkL07Model::create_link(const std::string& name, double bandwidth, double latency,
227 s4u::Link::SharingPolicy policy)
229 return new LinkL07(this, name, bandwidth, latency, policy);
236 CpuL07::CpuL07(CpuL07Model* model, simgrid::s4u::Host* host, const std::vector<double>& speed_per_pstate, int core)
237 : Cpu(model, host, model->get_maxmin_system()->constraint_new(this, speed_per_pstate.front()), speed_per_pstate,
242 CpuL07::~CpuL07()=default;
244 LinkL07::LinkL07(NetworkL07Model* model, const std::string& name, double bandwidth, double latency,
245 s4u::Link::SharingPolicy policy)
246 : LinkImpl(model, name, model->get_maxmin_system()->constraint_new(this, bandwidth))
248 bandwidth_.peak = bandwidth;
249 latency_.peak = latency;
251 if (policy == s4u::Link::SharingPolicy::FATPIPE)
252 get_constraint()->unshare();
254 s4u::Link::on_creation(this->piface_);
257 kernel::resource::CpuAction* CpuL07::execution_start(double size)
259 std::vector<s4u::Host*> host_list = {get_host()};
261 double* flops_amount = new double[host_list.size()]();
262 flops_amount[0] = size;
264 kernel::resource::CpuAction* res =
265 static_cast<CpuL07Model*>(get_model())->hostModel_->execute_parallel(host_list, flops_amount, nullptr, -1);
266 static_cast<L07Action*>(res)->free_arrays_ = true;
270 kernel::resource::CpuAction* CpuL07::sleep(double duration)
272 L07Action *action = static_cast<L07Action*>(execution_start(1.0));
273 action->set_max_duration(duration);
274 action->set_suspend_state(kernel::resource::Action::SuspendStates::SLEEPING);
275 get_model()->get_maxmin_system()->update_variable_weight(action->get_variable(), 0.0);
280 bool CpuL07::is_used()
282 return get_model()->get_maxmin_system()->constraint_used(get_constraint());
285 /** @brief take into account changes of speed (either load or max) */
286 void CpuL07::on_speed_change()
288 kernel::lmm::Variable* var = nullptr;
289 const kernel::lmm::Element* elem = nullptr;
291 get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), speed_.peak * speed_.scale);
292 while ((var = get_constraint()->get_variable(&elem))) {
293 kernel::resource::Action* action = static_cast<kernel::resource::Action*>(var->get_id());
295 get_model()->get_maxmin_system()->update_variable_bound(action->get_variable(), speed_.scale * speed_.peak);
298 Cpu::on_speed_change();
301 bool LinkL07::is_used()
303 return get_model()->get_maxmin_system()->constraint_used(get_constraint());
306 void CpuL07::apply_event(kernel::profile::Event* triggered, double value)
308 XBT_DEBUG("Updating cpu %s (%p) with value %g", get_cname(), this, value);
309 if (triggered == speed_.event) {
310 speed_.scale = value;
312 tmgr_trace_event_unref(&speed_.event);
314 } else if (triggered == state_event_) {
317 XBT_VERB("Restart processes on host %s", get_host()->get_cname());
318 get_host()->turn_on();
321 get_host()->turn_off();
322 tmgr_trace_event_unref(&state_event_);
325 xbt_die("Unknown event!\n");
329 void LinkL07::apply_event(kernel::profile::Event* 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()
381 delete[] computationAmount_;
382 delete[] communicationAmount_;
386 void L07Action::updateBound()
388 double lat_current = 0.0;
390 size_t host_count = hostList_.size();
392 if (communicationAmount_ != nullptr) {
393 for (size_t i = 0; i < host_count; i++) {
394 for (size_t j = 0; j < host_count; j++) {
396 if (communicationAmount_[i * host_count + j] > 0) {
398 std::vector<kernel::resource::LinkImpl*> route;
399 hostList_.at(i)->route_to(hostList_.at(j), route, &lat);
401 lat_current = std::max(lat_current, lat * communicationAmount_[i * host_count + j]);
406 double lat_bound = kernel::resource::NetworkModel::cfg_tcp_gamma / (2.0 * lat_current);
407 XBT_DEBUG("action (%p) : lat_bound = %g", this, lat_bound);
408 if ((latency_ <= 0.0) && is_running()) {
410 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), lat_bound);
412 get_model()->get_maxmin_system()->update_variable_bound(get_variable(), std::min(rate_, lat_bound));