src/surf/ptask_L07.cpp

   1 /* Copyright (c) 2007-2021. The SimGrid Team. All rights reserved.          */
   2
   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. */
   5
   6 #include "ptask_L07.hpp"
   7 #include "simgrid/kernel/routing/NetZoneImpl.hpp"
   8 #include "simgrid/s4u/Engine.hpp"
   9 #include "src/kernel/EngineImpl.hpp"
  10 #include "src/kernel/resource/profile/Event.hpp"
  11 #include "surf/surf.hpp"
  12 #include "xbt/config.hpp"
  13
  14 #include <unordered_set>
  15
  16 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(res_host);
  17 XBT_LOG_EXTERNAL_CATEGORY(xbt_cfg);
  18
  19 /**************************************/
  20 /*** Resource Creation & Destruction **/
  21 /**************************************/
  22 void surf_host_model_init_ptask_L07()
  23 {
  24   XBT_CINFO(xbt_cfg, "Switching to the L07 model to handle parallel tasks.");
  25
  26   auto host_model = std::make_shared<simgrid::surf::HostL07Model>();
  27   simgrid::kernel::EngineImpl::get_instance()->add_model(simgrid::kernel::resource::Model::Type::HOST, host_model,
  28                                                          true);
  29   simgrid::s4u::Engine::get_instance()->get_netzone_root()->get_impl()->set_host_model(host_model);
  30 }
  31
  32 namespace simgrid {
  33 namespace surf {
  34
  35 HostL07Model::HostL07Model() : HostModel()
  36 {
  37   auto* maxmin_system = new simgrid::kernel::lmm::FairBottleneck(true /* selective update */);
  38   set_maxmin_system(maxmin_system);
  39
  40   auto net_model = std::make_shared<NetworkL07Model>(this, maxmin_system);
  41   auto engine    = simgrid::kernel::EngineImpl::get_instance();
  42   engine->add_model(simgrid::kernel::resource::Model::Type::NETWORK, net_model, true);
  43   simgrid::s4u::Engine::get_instance()->get_netzone_root()->get_impl()->set_network_model(net_model);
  44
  45   auto cpu_model = std::make_shared<CpuL07Model>(this, maxmin_system);
  46   engine->add_model(simgrid::kernel::resource::Model::Type::CPU_PM, cpu_model, true);
  47   simgrid::s4u::Engine::get_instance()->get_netzone_root()->get_impl()->set_cpu_pm_model(cpu_model);
  48 }
  49
  50 CpuL07Model::CpuL07Model(HostL07Model* hmodel, kernel::lmm::System* sys) : hostModel_(hmodel)
  51 {
  52   set_maxmin_system(sys);
  53 }
  54
  55 CpuL07Model::~CpuL07Model()
  56 {
  57   set_maxmin_system(nullptr);
  58 }
  59
  60 NetworkL07Model::NetworkL07Model(HostL07Model* hmodel, kernel::lmm::System* sys) : hostModel_(hmodel)
  61 {
  62   set_maxmin_system(sys);
  63   loopback_ = NetworkL07Model::create_link(
  64                   "__loopback__", std::vector<double>{simgrid::config::get_value<double>("network/loopback-bw")},
  65                   s4u::Link::SharingPolicy::FATPIPE)
  66                   ->set_latency(simgrid::config::get_value<double>("network/loopback-lat"));
  67   loopback_->seal();
  68 }
  69
  70 NetworkL07Model::~NetworkL07Model()
  71 {
  72   set_maxmin_system(nullptr);
  73 }
  74
  75 double HostL07Model::next_occurring_event(double now)
  76 {
  77   double min = HostModel::next_occurring_event_full(now);
  78   for (kernel::resource::Action const& action : *get_started_action_set()) {
  79     const auto& net_action = static_cast<const L07Action&>(action);
  80     if (net_action.get_latency() > 0 && (min < 0 || net_action.get_latency() < min)) {
  81       min = net_action.get_latency();
  82       XBT_DEBUG("Updating min with %p (start %f): %f", &net_action, net_action.get_start_time(), min);
  83     }
  84   }
  85   XBT_DEBUG("min value: %f", min);
  86
  87   return min;
  88 }
  89
  90 void HostL07Model::update_actions_state(double /*now*/, double delta)
  91 {
  92   for (auto it = std::begin(*get_started_action_set()); it != std::end(*get_started_action_set());) {
  93     auto& action = static_cast<L07Action&>(*it);
  94     ++it; // increment iterator here since the following calls to action.finish() may invalidate it
  95     if (action.get_latency() > 0) {
  96       if (action.get_latency() > delta) {
  97         action.update_latency(delta, sg_surf_precision);
  98       } else {
  99         action.set_latency(0.0);
 100       }
 101       if ((action.get_latency() <= 0.0) && (action.is_suspended() == 0)) {
 102         action.updateBound();
 103         get_maxmin_system()->update_variable_penalty(action.get_variable(), 1.0);
 104         action.set_last_update();
 105       }
 106     }
 107     XBT_DEBUG("Action (%p) : remains (%g) updated by %g.", &action, action.get_remains(),
 108               action.get_variable()->get_value() * delta);
 109     action.update_remains(action.get_variable()->get_value() * delta);
 110     action.update_max_duration(delta);
 111
 112     XBT_DEBUG("Action (%p) : remains (%g).", &action, action.get_remains());
 113
 114     /* In the next if cascade, the action can be finished either because:
 115      *  - The amount of remaining work reached 0
 116      *  - The max duration was reached
 117      * If it's not done, it may have failed.
 118      */
 119
 120     if (((action.get_remains() <= 0) && (action.get_variable()->get_penalty() > 0)) ||
 121         ((action.get_max_duration() != NO_MAX_DURATION) && (action.get_max_duration() <= 0))) {
 122       action.finish(kernel::resource::Action::State::FINISHED);
 123       continue;
 124     }
 125
 126     /* Need to check that none of the model has failed */
 127     int i                               = 0;
 128     const kernel::lmm::Constraint* cnst = action.get_variable()->get_constraint(i);
 129     while (cnst != nullptr) {
 130       i++;
 131       const kernel::resource::Resource* constraint_id = cnst->get_id();
 132       if (not constraint_id->is_on()) {
 133         XBT_DEBUG("Action (%p) Failed!!", &action);
 134         action.finish(kernel::resource::Action::State::FAILED);
 135         break;
 136       }
 137       cnst = action.get_variable()->get_constraint(i);
 138     }
 139   }
 140 }
 141
 142 kernel::resource::CpuAction* HostL07Model::execute_parallel(const std::vector<s4u::Host*>& host_list,
 143                                                             const double* flops_amount, const double* bytes_amount,
 144                                                             double rate)
 145 {
 146   return new L07Action(this, host_list, flops_amount, bytes_amount, rate);
 147 }
 148
 149 L07Action::L07Action(kernel::resource::Model* model, const std::vector<s4u::Host*>& host_list,
 150                      const double* flops_amount, const double* bytes_amount, double rate)
 151     : CpuAction(model, 1.0, false), computationAmount_(flops_amount), communicationAmount_(bytes_amount), rate_(rate)
 152 {
 153   size_t link_nb      = 0;
 154   size_t used_host_nb = 0; /* Only the hosts with something to compute (>0 flops) are counted) */
 155   double latency      = 0.0;
 156   this->set_last_update();
 157
 158   hostList_.insert(hostList_.end(), host_list.begin(), host_list.end());
 159
 160   if (flops_amount != nullptr)
 161     used_host_nb += std::count_if(flops_amount, flops_amount + host_list.size(), [](double x) { return x > 0.0; });
 162
 163   /* Compute the number of affected resources... */
 164   if (bytes_amount != nullptr) {
 165     std::unordered_set<const char*> affected_links;
 166
 167     for (size_t k = 0; k < host_list.size() * host_list.size(); k++) {
 168       if (bytes_amount[k] <= 0)
 169         continue;
 170
 171       double lat = 0.0;
 172       std::vector<kernel::resource::LinkImpl*> route;
 173       hostList_[k / host_list.size()]->route_to(hostList_[k % host_list.size()], route, &lat);
 174       latency = std::max(latency, lat);
 175
 176       for (auto const& link : route)
 177         affected_links.insert(link->get_cname());
 178     }
 179
 180     link_nb = affected_links.size();
 181   }
 182
 183   XBT_DEBUG("Creating a parallel task (%p) with %zu hosts and %zu unique links.", this, host_list.size(), link_nb);
 184   latency_ = latency;
 185
 186   set_variable(
 187       model->get_maxmin_system()->variable_new(this, 1.0, (rate > 0 ? rate : -1.0), host_list.size() + link_nb));
 188
 189   if (latency_ > 0)
 190     model->get_maxmin_system()->update_variable_penalty(get_variable(), 0.0);
 191
 192   /* Expand it for the CPUs even if there is nothing to compute, to make sure that it gets expended even if there is no
 193    * communication either */
 194   double bound = std::numeric_limits<double>::max();
 195   for (size_t i = 0; i < host_list.size(); i++) {
 196     model->get_maxmin_system()->expand(host_list[i]->pimpl_cpu->get_constraint(), get_variable(),
 197                                        (flops_amount == nullptr ? 0.0 : flops_amount[i]));
 198     if (flops_amount && flops_amount[i] > 0)
 199       bound = std::min(bound, host_list[i]->pimpl_cpu->get_speed(1.0) * host_list[i]->pimpl_cpu->get_speed_ratio() /
 200                                   flops_amount[i]);
 201   }
 202   if (bound < std::numeric_limits<double>::max())
 203     model->get_maxmin_system()->update_variable_bound(get_variable(), bound);
 204
 205   if (bytes_amount != nullptr) {
 206     for (size_t k = 0; k < host_list.size() * host_list.size(); k++) {
 207       if (bytes_amount[k] <= 0.0)
 208         continue;
 209       std::vector<kernel::resource::LinkImpl*> route;
 210       hostList_[k / host_list.size()]->route_to(hostList_[k % host_list.size()], route, nullptr);
 211
 212       for (auto const& link : route)
 213         model->get_maxmin_system()->expand_add(link->get_constraint(), this->get_variable(), bytes_amount[k]);
 214     }
 215   }
 216
 217   if (link_nb + used_host_nb == 0) {
 218     this->set_cost(1.0);
 219     this->set_remains(0.0);
 220   }
 221 }
 222
 223 kernel::resource::Action* NetworkL07Model::communicate(s4u::Host* src, s4u::Host* dst, double size, double rate)
 224 {
 225   std::vector<s4u::Host*> host_list = {src, dst};
 226   const auto* flops_amount          = new double[2]();
 227   auto* bytes_amount                = new double[4]();
 228
 229   bytes_amount[1] = size;
 230
 231   kernel::resource::Action* res = hostModel_->execute_parallel(host_list, flops_amount, bytes_amount, rate);
 232   static_cast<L07Action*>(res)->free_arrays_ = true;
 233   return res;
 234 }
 235
 236 kernel::resource::Cpu* CpuL07Model::create_cpu(s4u::Host* host, const std::vector<double>& speed_per_pstate)
 237 {
 238   return (new CpuL07(host, speed_per_pstate))->set_model(this);
 239 }
 240
 241 kernel::resource::LinkImpl* NetworkL07Model::create_link(const std::string& name, const std::vector<double>& bandwidths,
 242                                                          s4u::Link::SharingPolicy policy)
 243 {
 244   xbt_assert(bandwidths.size() == 1, "Non WIFI link must have only 1 bandwidth.");
 245   auto link = new LinkL07(name, bandwidths[0], policy, get_maxmin_system());
 246   link->set_model(this);
 247   return link;
 248 }
 249
 250 /************
 251  * Resource *
 252  ************/
 253
 254 kernel::resource::CpuAction* CpuL07::execution_start(double size)
 255 {
 256   std::vector<s4u::Host*> host_list = {get_iface()};
 257
 258   auto* flops_amount = new double[host_list.size()]();
 259   flops_amount[0]    = size;
 260
 261   kernel::resource::CpuAction* res =
 262       static_cast<CpuL07Model*>(get_model())->hostModel_->execute_parallel(host_list, flops_amount, nullptr, -1);
 263   static_cast<L07Action*>(res)->free_arrays_ = true;
 264   return res;
 265 }
 266
 267 kernel::resource::CpuAction* CpuL07::sleep(double duration)
 268 {
 269   auto* action = static_cast<L07Action*>(execution_start(1.0));
 270   action->set_max_duration(duration);
 271   action->set_suspend_state(kernel::resource::Action::SuspendStates::SLEEPING);
 272   get_model()->get_maxmin_system()->update_variable_penalty(action->get_variable(), 0.0);
 273
 274   return action;
 275 }
 276
 277 bool CpuL07::is_used() const
 278 {
 279   return get_model()->get_maxmin_system()->constraint_used(get_constraint());
 280 }
 281
 282 /** @brief take into account changes of speed (either load or max) */
 283 void CpuL07::on_speed_change()
 284 {
 285   const kernel::lmm::Variable* var;
 286   const kernel::lmm::Element* elem = nullptr;
 287
 288   get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), speed_.peak * speed_.scale);
 289   while ((var = get_constraint()->get_variable(&elem))) {
 290     const kernel::resource::Action* action = var->get_id();
 291
 292     get_model()->get_maxmin_system()->update_variable_bound(action->get_variable(), speed_.scale * speed_.peak);
 293   }
 294
 295   Cpu::on_speed_change();
 296 }
 297
 298 LinkL07::LinkL07(const std::string& name, double bandwidth, s4u::Link::SharingPolicy policy,
 299                  kernel::lmm::System* system)
 300     : LinkImpl(name)
 301 {
 302   this->set_constraint(system->constraint_new(this, bandwidth));
 303   bandwidth_.peak = bandwidth;
 304
 305   if (policy == s4u::Link::SharingPolicy::FATPIPE)
 306     get_constraint()->unshare();
 307 }
 308
 309 bool LinkL07::is_used() const
 310 {
 311   return get_model()->get_maxmin_system()->constraint_used(get_constraint());
 312 }
 313
 314 void CpuL07::apply_event(kernel::profile::Event* triggered, double value)
 315 {
 316   XBT_DEBUG("Updating cpu %s (%p) with value %g", get_cname(), this, value);
 317   if (triggered == speed_.event) {
 318     speed_.scale = value;
 319     on_speed_change();
 320     tmgr_trace_event_unref(&speed_.event);
 321
 322   } else if (triggered == state_event_) {
 323     if (value > 0) {
 324       if (not is_on()) {
 325         XBT_VERB("Restart actors on host %s", get_iface()->get_cname());
 326         get_iface()->turn_on();
 327       }
 328     } else
 329       get_iface()->turn_off();
 330     tmgr_trace_event_unref(&state_event_);
 331
 332   } else {
 333     xbt_die("Unknown event!\n");
 334   }
 335 }
 336
 337 void LinkL07::apply_event(kernel::profile::Event* triggered, double value)
 338 {
 339   XBT_DEBUG("Updating link %s (%p) with value=%f", get_cname(), this, value);
 340   if (triggered == bandwidth_.event) {
 341     set_bandwidth(value);
 342     tmgr_trace_event_unref(&bandwidth_.event);
 343
 344   } else if (triggered == latency_.event) {
 345     set_latency(value);
 346     tmgr_trace_event_unref(&latency_.event);
 347
 348   } else if (triggered == state_event_) {
 349     if (value > 0)
 350       turn_on();
 351     else
 352       turn_off();
 353     tmgr_trace_event_unref(&state_event_);
 354
 355   } else {
 356     xbt_die("Unknown event ! \n");
 357   }
 358 }
 359
 360 void LinkL07::set_bandwidth(double value)
 361 {
 362   bandwidth_.peak = value;
 363   LinkImpl::on_bandwidth_change();
 364
 365   get_model()->get_maxmin_system()->update_constraint_bound(get_constraint(), bandwidth_.peak * bandwidth_.scale);
 366 }
 367
 368 kernel::resource::LinkImpl* LinkL07::set_latency(double value)
 369 {
 370   latency_check(value);
 371   const kernel::lmm::Variable* var;
 372   L07Action* action;
 373   const kernel::lmm::Element* elem = nullptr;
 374
 375   latency_.peak = value;
 376   while ((var = get_constraint()->get_variable(&elem))) {
 377     action = static_cast<L07Action*>(var->get_id());
 378     action->updateBound();
 379   }
 380   return this;
 381 }
 382 LinkL07::~LinkL07() = default;
 383
 384 /**********
 385  * Action *
 386  **********/
 387
 388 L07Action::~L07Action()
 389 {
 390   if (free_arrays_) {
 391     delete[] computationAmount_;
 392     delete[] communicationAmount_;
 393   }
 394 }
 395
 396 void L07Action::updateBound()
 397 {
 398   double lat_current = 0.0;
 399
 400   size_t host_count = hostList_.size();
 401
 402   if (communicationAmount_ != nullptr) {
 403     for (size_t i = 0; i < host_count; i++) {
 404       for (size_t j = 0; j < host_count; j++) {
 405         if (communicationAmount_[i * host_count + j] > 0) {
 406           double lat = 0.0;
 407           std::vector<kernel::resource::LinkImpl*> route;
 408           hostList_.at(i)->route_to(hostList_.at(j), route, &lat);
 409
 410           lat_current = std::max(lat_current, lat * communicationAmount_[i * host_count + j]);
 411         }
 412       }
 413     }
 414   }
 415   double lat_bound = kernel::resource::NetworkModel::cfg_tcp_gamma / (2.0 * lat_current);
 416   XBT_DEBUG("action (%p) : lat_bound = %g", this, lat_bound);
 417   if ((latency_ <= 0.0) && is_running()) {
 418     if (rate_ < 0)
 419       get_model()->get_maxmin_system()->update_variable_bound(get_variable(), lat_bound);
 420     else
 421       get_model()->get_maxmin_system()->update_variable_bound(get_variable(), std::min(rate_, lat_bound));
 422   }
 423 }
 424
 425 } // namespace surf
 426 } // namespace simgrid