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