Logo AND Algorithmique Numérique Distribuée

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