Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
Use const& for std::string parameter (sonar, again).
[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 "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>("Host_Ptask");
27   simgrid::kernel::EngineImpl::get_instance()->add_model(host_model);
28   simgrid::s4u::Engine::get_instance()->get_netzone_root()->get_impl()->set_host_model(host_model);
29 }
30
31 namespace simgrid {
32 namespace surf {
33
34 HostL07Model::HostL07Model(const std::string& name) : HostModel(name)
35 {
36   auto* maxmin_system = new simgrid::kernel::lmm::FairBottleneck(true /* selective update */);
37   set_maxmin_system(maxmin_system);
38
39   auto net_model = std::make_shared<NetworkL07Model>("Network_Ptask", this, maxmin_system);
40   auto engine    = simgrid::kernel::EngineImpl::get_instance();
41   engine->add_model(net_model);
42   simgrid::s4u::Engine::get_instance()->get_netzone_root()->get_impl()->set_network_model(net_model);
43
44   auto cpu_model = std::make_shared<CpuL07Model>("Cpu_Ptask", this, maxmin_system);
45   engine->add_model(cpu_model);
46   simgrid::s4u::Engine::get_instance()->get_netzone_root()->get_impl()->set_cpu_pm_model(cpu_model);
47 }
48
49 CpuL07Model::CpuL07Model(const std::string& name, HostL07Model* hmodel, kernel::lmm::System* sys)
50     : CpuModel(name), hostModel_(hmodel)
51 {
52   set_maxmin_system(sys);
53 }
54
55 CpuL07Model::~CpuL07Model()
56 {
57   set_maxmin_system(nullptr);
58 }
59
60 NetworkL07Model::NetworkL07Model(const std::string& name, HostL07Model* hmodel, kernel::lmm::System* sys)
61     : NetworkModel(name), hostModel_(hmodel)
62 {
63   set_maxmin_system(sys);
64   loopback_ = NetworkL07Model::create_link(
65                   "__loopback__", std::vector<double>{simgrid::config::get_value<double>("network/loopback-bw")},
66                   s4u::Link::SharingPolicy::FATPIPE)
67                   ->set_latency(simgrid::config::get_value<double>("network/loopback-lat"));
68   loopback_->seal();
69 }
70
71 NetworkL07Model::~NetworkL07Model()
72 {
73   set_maxmin_system(nullptr);
74 }
75
76 double HostL07Model::next_occurring_event(double now)
77 {
78   double min = HostModel::next_occurring_event_full(now);
79   for (kernel::resource::Action const& action : *get_started_action_set()) {
80     const auto& net_action = static_cast<const L07Action&>(action);
81     if (net_action.get_latency() > 0 && (min < 0 || net_action.get_latency() < min)) {
82       min = net_action.get_latency();
83       XBT_DEBUG("Updating min with %p (start %f): %f", &net_action, net_action.get_start_time(), min);
84     }
85   }
86   XBT_DEBUG("min value: %f", min);
87
88   return min;
89 }
90
91 void HostL07Model::update_actions_state(double /*now*/, double delta)
92 {
93   for (auto it = std::begin(*get_started_action_set()); it != std::end(*get_started_action_set());) {
94     auto& action = static_cast<L07Action&>(*it);
95     ++it; // increment iterator here since the following calls to action.finish() may invalidate it
96     if (action.get_latency() > 0) {
97       if (action.get_latency() > delta) {
98         action.update_latency(delta, sg_surf_precision);
99       } else {
100         action.set_latency(0.0);
101       }
102       if ((action.get_latency() <= 0.0) && (action.is_suspended() == 0)) {
103         action.updateBound();
104         get_maxmin_system()->update_variable_penalty(action.get_variable(), 1.0);
105         action.set_last_update();
106       }
107     }
108     XBT_DEBUG("Action (%p) : remains (%g) updated by %g.", &action, action.get_remains(),
109               action.get_variable()->get_value() * delta);
110     action.update_remains(action.get_variable()->get_value() * delta);
111     action.update_max_duration(delta);
112
113     XBT_DEBUG("Action (%p) : remains (%g).", &action, action.get_remains());
114
115     /* In the next if cascade, the action can be finished either because:
116      *  - The amount of remaining work reached 0
117      *  - The max duration was reached
118      * If it's not done, it may have failed.
119      */
120
121     if (((action.get_remains() <= 0) && (action.get_variable()->get_penalty() > 0)) ||
122         ((action.get_max_duration() != NO_MAX_DURATION) && (action.get_max_duration() <= 0))) {
123       action.finish(kernel::resource::Action::State::FINISHED);
124       continue;
125     }
126
127     /* Need to check that none of the model has failed */
128     int i                               = 0;
129     const kernel::lmm::Constraint* cnst = action.get_variable()->get_constraint(i);
130     while (cnst != nullptr) {
131       i++;
132       const kernel::resource::Resource* constraint_id = cnst->get_id();
133       if (not constraint_id->is_on()) {
134         XBT_DEBUG("Action (%p) Failed!!", &action);
135         action.finish(kernel::resource::Action::State::FAILED);
136         break;
137       }
138       cnst = action.get_variable()->get_constraint(i);
139     }
140   }
141 }
142
143 kernel::resource::CpuAction* HostL07Model::execute_parallel(const std::vector<s4u::Host*>& host_list,
144                                                             const double* flops_amount, const double* bytes_amount,
145                                                             double rate)
146 {
147   return new L07Action(this, host_list, flops_amount, bytes_amount, rate);
148 }
149
150 L07Action::L07Action(kernel::resource::Model* model, const std::vector<s4u::Host*>& host_list,
151                      const double* flops_amount, const double* bytes_amount, double rate)
152     : CpuAction(model, 1.0, false), computationAmount_(flops_amount), communicationAmount_(bytes_amount), rate_(rate)
153 {
154   size_t link_nb      = 0;
155   size_t used_host_nb = 0; /* Only the hosts with something to compute (>0 flops) are counted) */
156   double latency      = 0.0;
157   this->set_last_update();
158
159   hostList_.insert(hostList_.end(), host_list.begin(), host_list.end());
160
161   if (flops_amount != nullptr)
162     used_host_nb += std::count_if(flops_amount, flops_amount + host_list.size(), [](double x) { return x > 0.0; });
163
164   /* Compute the number of affected resources... */
165   if (bytes_amount != nullptr) {
166     std::unordered_set<const char*> affected_links;
167
168     for (size_t k = 0; k < host_list.size() * host_list.size(); k++) {
169       if (bytes_amount[k] <= 0)
170         continue;
171
172       double lat = 0.0;
173       std::vector<kernel::resource::LinkImpl*> route;
174       hostList_[k / host_list.size()]->route_to(hostList_[k % host_list.size()], route, &lat);
175       latency = std::max(latency, lat);
176
177       for (auto const& link : route)
178         affected_links.insert(link->get_cname());
179     }
180
181     link_nb = affected_links.size();
182   }
183
184   XBT_DEBUG("Creating a parallel task (%p) with %zu hosts and %zu unique links.", this, host_list.size(), link_nb);
185   latency_ = latency;
186
187   set_variable(
188       model->get_maxmin_system()->variable_new(this, 1.0, (rate > 0 ? rate : -1.0), host_list.size() + link_nb));
189
190   if (latency_ > 0)
191     model->get_maxmin_system()->update_variable_penalty(get_variable(), 0.0);
192
193   /* Expand it for the CPUs even if there is nothing to compute, to make sure that it gets expended even if there is no
194    * communication either */
195   double bound = std::numeric_limits<double>::max();
196   for (size_t i = 0; i < host_list.size(); i++) {
197     model->get_maxmin_system()->expand(host_list[i]->pimpl_cpu->get_constraint(), get_variable(),
198                                        (flops_amount == nullptr ? 0.0 : flops_amount[i]));
199     if (flops_amount && flops_amount[i] > 0)
200       bound = std::min(bound, host_list[i]->pimpl_cpu->get_speed(1.0) * host_list[i]->pimpl_cpu->get_speed_ratio() /
201                                   flops_amount[i]);
202   }
203   if (bound < std::numeric_limits<double>::max())
204     model->get_maxmin_system()->update_variable_bound(get_variable(), bound);
205
206   if (bytes_amount != nullptr) {
207     for (size_t k = 0; k < host_list.size() * host_list.size(); k++) {
208       if (bytes_amount[k] <= 0.0)
209         continue;
210       std::vector<kernel::resource::LinkImpl*> route;
211       hostList_[k / host_list.size()]->route_to(hostList_[k % host_list.size()], route, nullptr);
212
213       for (auto const& link : route)
214         model->get_maxmin_system()->expand_add(link->get_constraint(), this->get_variable(), bytes_amount[k]);
215     }
216   }
217
218   if (link_nb + used_host_nb == 0) {
219     this->set_cost(1.0);
220     this->set_remains(0.0);
221   }
222 }
223
224 kernel::resource::Action* NetworkL07Model::communicate(s4u::Host* src, s4u::Host* dst, double size, double rate)
225 {
226   std::vector<s4u::Host*> host_list = {src, dst};
227   const auto* flops_amount          = new double[2]();
228   auto* bytes_amount                = new double[4]();
229
230   bytes_amount[1] = size;
231
232   kernel::resource::Action* res = hostModel_->execute_parallel(host_list, flops_amount, bytes_amount, rate);
233   static_cast<L07Action*>(res)->free_arrays_ = true;
234   return res;
235 }
236
237 kernel::resource::Cpu* CpuL07Model::create_cpu(s4u::Host* host, const std::vector<double>& speed_per_pstate)
238 {
239   return (new CpuL07(host, speed_per_pstate))->set_model(this);
240 }
241
242 kernel::resource::LinkImpl* NetworkL07Model::create_link(const std::string& name, const std::vector<double>& bandwidths,
243                                                          s4u::Link::SharingPolicy policy)
244 {
245   xbt_assert(bandwidths.size() == 1, "Non WIFI link must have only 1 bandwidth.");
246   auto link = new LinkL07(name, bandwidths[0], policy, get_maxmin_system());
247   link->set_model(this);
248   return link;
249 }
250
251 /************
252  * Resource *
253  ************/
254
255 kernel::resource::CpuAction* CpuL07::execution_start(double size)
256 {
257   std::vector<s4u::Host*> host_list = {get_iface()};
258
259   auto* flops_amount = new double[host_list.size()]();
260   flops_amount[0]    = size;
261
262   kernel::resource::CpuAction* res =
263       static_cast<CpuL07Model*>(get_model())->hostModel_->execute_parallel(host_list, flops_amount, nullptr, -1);
264   static_cast<L07Action*>(res)->free_arrays_ = true;
265   return res;
266 }
267
268 kernel::resource::CpuAction* CpuL07::sleep(double duration)
269 {
270   auto* action = static_cast<L07Action*>(execution_start(1.0));
271   action->set_max_duration(duration);
272   action->set_suspend_state(kernel::resource::Action::SuspendStates::SLEEPING);
273   get_model()->get_maxmin_system()->update_variable_penalty(action->get_variable(), 0.0);
274
275   return action;
276 }
277
278 bool CpuL07::is_used() const
279 {
280   return get_model()->get_maxmin_system()->constraint_used(get_constraint());
281 }
282
283 /** @brief take into account changes of speed (either load or max) */
284 void CpuL07::on_speed_change()
285 {
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 (const auto* 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::Element* elem = nullptr;
372
373   latency_.peak = value;
374   while (const auto* var = get_constraint()->get_variable(&elem)) {
375     auto* action = static_cast<L07Action*>(var->get_id());
376     action->updateBound();
377   }
378   return this;
379 }
380 LinkL07::~LinkL07() = default;
381
382 /**********
383  * Action *
384  **********/
385
386 L07Action::~L07Action()
387 {
388   if (free_arrays_) {
389     delete[] computationAmount_;
390     delete[] communicationAmount_;
391   }
392 }
393
394 void L07Action::updateBound()
395 {
396   double lat_current = 0.0;
397
398   size_t host_count = hostList_.size();
399
400   if (communicationAmount_ != nullptr) {
401     for (size_t i = 0; i < host_count; i++) {
402       for (size_t j = 0; j < host_count; j++) {
403         if (communicationAmount_[i * host_count + j] > 0) {
404           double lat = 0.0;
405           std::vector<kernel::resource::LinkImpl*> route;
406           hostList_.at(i)->route_to(hostList_.at(j), route, &lat);
407
408           lat_current = std::max(lat_current, lat * communicationAmount_[i * host_count + j]);
409         }
410       }
411     }
412   }
413   double lat_bound = kernel::resource::NetworkModel::cfg_tcp_gamma / (2.0 * lat_current);
414   XBT_DEBUG("action (%p) : lat_bound = %g", this, lat_bound);
415   if ((latency_ <= 0.0) && is_running()) {
416     if (rate_ < 0)
417       get_model()->get_maxmin_system()->update_variable_bound(get_variable(), lat_bound);
418     else
419       get_model()->get_maxmin_system()->update_variable_bound(get_variable(), std::min(rate_, lat_bound));
420   }
421 }
422
423 } // namespace surf
424 } // namespace simgrid