1 /* Copyright (c) 2007-2010, 2013-2015. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
10 #include <unordered_set>
12 #include "ptask_L07.hpp"
14 #include "cpu_interface.hpp"
15 #include "surf_routing.hpp"
18 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(surf_host);
19 XBT_LOG_EXTERNAL_CATEGORY(xbt_cfg);
21 /**************************************/
22 /*** Resource Creation & Destruction **/
23 /**************************************/
24 void surf_host_model_init_ptask_L07()
26 XBT_CINFO(xbt_cfg,"Switching to the L07 model to handle parallel tasks.");
27 xbt_assert(!surf_cpu_model_pm, "CPU model type already defined");
28 xbt_assert(!surf_network_model, "network model type already defined");
30 surf_host_model = new simgrid::surf::HostL07Model();
31 all_existing_models->push_back(surf_host_model);
38 HostL07Model::HostL07Model() : HostModel() {
39 maxminSystem_ = lmm_system_new(true /* lazy */);
40 maxminSystem_->solve_fun = &bottleneck_solve;
41 surf_network_model = new NetworkL07Model(this,maxminSystem_);
42 surf_cpu_model_pm = new CpuL07Model(this,maxminSystem_);
45 HostL07Model::~HostL07Model() = default;
47 CpuL07Model::CpuL07Model(HostL07Model *hmodel,lmm_system_t sys)
53 CpuL07Model::~CpuL07Model() {
54 lmm_system_free(maxminSystem_);
55 maxminSystem_ = nullptr;
57 NetworkL07Model::NetworkL07Model(HostL07Model *hmodel, lmm_system_t sys)
62 loopback_ = createLink("__loopback__", 498000000, 0.000015, SURF_LINK_FATPIPE);
64 NetworkL07Model::~NetworkL07Model()
66 maxminSystem_ = nullptr; // Avoid multi-free
70 double HostL07Model::nextOccuringEvent(double now)
72 double min = HostModel::nextOccuringEventFull(now);
73 for (auto it(getRunningActionSet()->begin()), itend(getRunningActionSet()->end()); it != itend ; ++it) {
74 L07Action *action = static_cast<L07Action*>(&*it);
75 if (action->latency_ > 0 && (min < 0 || action->latency_ < min)) {
76 min = action->latency_;
77 XBT_DEBUG("Updating min with %p (start %f): %f", action, action->getStartTime(), min);
80 XBT_DEBUG("min value: %f", min);
85 void HostL07Model::updateActionsState(double /*now*/, double delta) {
88 ActionList *actionSet = getRunningActionSet();
90 for(ActionList::iterator it = actionSet->begin(), itNext = it
91 ; it != actionSet->end()
94 action = static_cast<L07Action*>(&*it);
95 if (action->latency_ > 0) {
96 if (action->latency_ > delta) {
97 double_update(&(action->latency_), delta, sg_surf_precision);
99 action->latency_ = 0.0;
101 if ((action->latency_ == 0.0) && (action->isSuspended() == 0)) {
102 action->updateBound();
103 lmm_update_variable_weight(maxminSystem_, action->getVariable(), 1.0);
106 XBT_DEBUG("Action (%p) : remains (%g) updated by %g.",
107 action, action->getRemains(), lmm_variable_getvalue(action->getVariable()) * delta);
108 action->updateRemains(lmm_variable_getvalue(action->getVariable()) * delta);
110 if (action->getMaxDuration() != NO_MAX_DURATION)
111 action->updateMaxDuration(delta);
113 XBT_DEBUG("Action (%p) : remains (%g).", action, action->getRemains());
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.
121 if ((action->getRemains() <= 0) &&
122 (lmm_get_variable_weight(action->getVariable()) > 0)) {
124 action->setState(Action::State::done);
125 } else if ((action->getMaxDuration() != NO_MAX_DURATION) &&
126 (action->getMaxDuration() <= 0)) {
128 action->setState(Action::State::done);
130 /* Need to check that none of the model has failed */
131 lmm_constraint_t cnst = nullptr;
134 while ((cnst = lmm_get_cnst_from_var(maxminSystem_, action->getVariable(), i++))) {
135 void *constraint_id = lmm_constraint_id(cnst);
136 if (static_cast<simgrid::surf::Resource*>(constraint_id)->isOff()) {
137 XBT_DEBUG("Action (%p) Failed!!", action);
139 action->setState(Action::State::failed);
148 Action *HostL07Model::executeParallelTask(int host_nb, sg_host_t *host_list,
149 double *flops_amount, double *bytes_amount,double rate) {
150 return new L07Action(this, host_nb, host_list, flops_amount, bytes_amount, rate);
153 L07Action::L07Action(Model *model, int host_nb, sg_host_t *host_list,
154 double *flops_amount, double *bytes_amount, double rate)
155 : CpuAction(model, 1, 0)
156 , computationAmount_(flops_amount)
157 , communicationAmount_(bytes_amount)
161 int nb_used_host = 0; /* Only the hosts with something to compute (>0 flops) are counted) */
162 double latency = 0.0;
164 this->hostList_->reserve(host_nb);
165 for (int i = 0; i<host_nb; i++)
166 this->hostList_->push_back(host_list[i]);
168 /* Compute the number of affected resources... */
169 if(bytes_amount != nullptr) {
170 std::unordered_set<const char*> affected_links;
172 for (int i = 0; i < host_nb; i++) {
173 for (int j = 0; j < host_nb; j++) {
175 if (bytes_amount[i * host_nb + j] > 0) {
178 std::vector<Link*> route;
179 hostList_->at(i)->routeTo(hostList_->at(j), &route, &lat);
180 latency = MAX(latency, lat);
182 for (auto link : route)
183 affected_links.insert(link->getName());
188 nb_link = affected_links.size();
191 for (int i = 0; i < host_nb; i++)
192 if (flops_amount[i] > 0)
195 XBT_DEBUG("Creating a parallel task (%p) with %d hosts and %d unique links.", this, host_nb, nb_link);
196 this->latency_ = latency;
198 this->variable_ = lmm_variable_new(model->getMaxminSystem(), this, 1.0,
199 (rate > 0 ? rate : -1.0),
202 if (this->latency_ > 0)
203 lmm_update_variable_weight(model->getMaxminSystem(), this->getVariable(), 0.0);
205 for (int i = 0; i < host_nb; i++)
206 lmm_expand(model->getMaxminSystem(), host_list[i]->pimpl_cpu->getConstraint(),
207 this->getVariable(), flops_amount[i]);
209 if(bytes_amount != nullptr) {
210 for (int i = 0; i < host_nb; i++) {
211 for (int j = 0; j < host_nb; j++) {
213 if (bytes_amount[i * host_nb + j] == 0.0)
216 std::vector<Link*> route;
217 hostList_->at(i)->routeTo(hostList_->at(j), &route, nullptr);
219 for (auto link : route)
220 lmm_expand_add(model->getMaxminSystem(), link->getConstraint(), this->getVariable(), bytes_amount[i * host_nb + j]);
225 if (nb_link + nb_used_host == 0) {
227 this->setRemains(0.0);
232 Action* NetworkL07Model::communicate(s4u::Host* src, s4u::Host* dst, double size, double rate)
234 sg_host_t*host_list = xbt_new0(sg_host_t, 2);
235 double *flops_amount = xbt_new0(double, 2);
236 double *bytes_amount = xbt_new0(double, 4);
240 bytes_amount[1] = size;
242 return hostModel_->executeParallelTask(2, host_list, flops_amount, bytes_amount, rate);
245 Cpu *CpuL07Model::createCpu(simgrid::s4u::Host *host, std::vector<double> *speedPerPstate, int core)
247 return new CpuL07(this, host, speedPerPstate, core);
250 Link* NetworkL07Model::createLink(const char* name, double bandwidth, double latency,
251 e_surf_link_sharing_policy_t policy)
253 return new LinkL07(this, name, bandwidth, latency, policy);
260 CpuL07::CpuL07(CpuL07Model* model, simgrid::s4u::Host* host, std::vector<double>* speedPerPstate, int core)
261 : Cpu(model, host, lmm_constraint_new(model->getMaxminSystem(), this, speedPerPstate->front()), speedPerPstate,
266 CpuL07::~CpuL07()=default;
268 LinkL07::LinkL07(NetworkL07Model* model, const char* name, double bandwidth, double latency,
269 e_surf_link_sharing_policy_t policy)
270 : Link(model, name, lmm_constraint_new(model->getMaxminSystem(), this, bandwidth))
272 bandwidth_.peak = bandwidth;
273 latency_.peak = latency;
275 if (policy == SURF_LINK_FATPIPE)
276 lmm_constraint_shared(getConstraint());
278 Link::onCreation(this);
281 Action *CpuL07::execution_start(double size)
283 sg_host_t*host_list = xbt_new0(sg_host_t, 1);
284 double *flops_amount = xbt_new0(double, 1);
286 host_list[0] = getHost();
287 flops_amount[0] = size;
289 return static_cast<CpuL07Model*>(getModel())->hostModel_->executeParallelTask(1, host_list, flops_amount, nullptr, -1);
292 Action *CpuL07::sleep(double duration)
294 L07Action *action = static_cast<L07Action*>(execution_start(1.0));
295 action->maxDuration_ = duration;
296 action->suspended_ = 2;
297 lmm_update_variable_weight(getModel()->getMaxminSystem(), action->getVariable(), 0.0);
302 bool CpuL07::isUsed(){
303 return lmm_constraint_used(getModel()->getMaxminSystem(), getConstraint());
306 /** @brief take into account changes of speed (either load or max) */
307 void CpuL07::onSpeedChange() {
308 lmm_variable_t var = nullptr;
309 lmm_element_t elem = nullptr;
311 lmm_update_constraint_bound(getModel()->getMaxminSystem(), getConstraint(), speed_.peak * speed_.scale);
312 while ((var = lmm_get_var_from_cnst (getModel()->getMaxminSystem(), getConstraint(), &elem))) {
313 Action *action = static_cast<Action*>(lmm_variable_id(var));
315 lmm_update_variable_bound(getModel()->getMaxminSystem(), action->getVariable(), speed_.scale * speed_.peak);
318 Cpu::onSpeedChange();
322 bool LinkL07::isUsed(){
323 return lmm_constraint_used(getModel()->getMaxminSystem(), getConstraint());
326 void CpuL07::apply_event(tmgr_trace_iterator_t triggered, double value){
327 XBT_DEBUG("Updating cpu %s (%p) with value %g", getName(), this, value);
328 if (triggered == speed_.event) {
329 speed_.scale = value;
331 tmgr_trace_event_unref(&speed_.event);
333 } else if (triggered == stateEvent_) {
338 tmgr_trace_event_unref(&stateEvent_);
341 xbt_die("Unknown event!\n");
345 void LinkL07::apply_event(tmgr_trace_iterator_t triggered, double value) {
346 XBT_DEBUG("Updating link %s (%p) with value=%f", getName(), this, value);
347 if (triggered == bandwidth_.event) {
349 tmgr_trace_event_unref(&bandwidth_.event);
351 } else if (triggered == latency_.event) {
353 tmgr_trace_event_unref(&latency_.event);
355 } else if (triggered == stateEvent_) {
360 tmgr_trace_event_unref(&stateEvent_);
363 xbt_die("Unknown event ! \n");
367 void LinkL07::setBandwidth(double value)
369 bandwidth_.peak = value;
370 lmm_update_constraint_bound(getModel()->getMaxminSystem(), getConstraint(), bandwidth_.peak * bandwidth_.scale);
373 void LinkL07::setLatency(double value)
375 lmm_variable_t var = nullptr;
377 lmm_element_t elem = nullptr;
379 latency_.peak = value;
380 while ((var = lmm_get_var_from_cnst(getModel()->getMaxminSystem(), getConstraint(), &elem))) {
381 action = static_cast<L07Action*>(lmm_variable_id(var));
382 action->updateBound();
390 L07Action::~L07Action(){
392 free(communicationAmount_);
393 free(computationAmount_);
396 void L07Action::updateBound()
398 double lat_current = 0.0;
399 double lat_bound = -1.0;
402 int hostNb = hostList_->size();
404 if (communicationAmount_ != nullptr) {
405 for (i = 0; i < hostNb; i++) {
406 for (j = 0; j < hostNb; j++) {
408 if (communicationAmount_[i * hostNb + j] > 0) {
410 std::vector<Link*> route;
411 hostList_->at(i)->routeTo(hostList_->at(j), &route, &lat);
413 lat_current = MAX(lat_current, lat * communicationAmount_[i * hostNb + j]);
418 lat_bound = sg_tcp_gamma / (2.0 * lat_current);
419 XBT_DEBUG("action (%p) : lat_bound = %g", this, lat_bound);
420 if ((latency_ == 0.0) && (suspended_ == 0)) {
422 lmm_update_variable_bound(getModel()->getMaxminSystem(), getVariable(), lat_bound);
424 lmm_update_variable_bound(getModel()->getMaxminSystem(), getVariable(), std::min(rate_, lat_bound));
428 int L07Action::unref()
432 if (action_hook.is_linked())
433 stateSet_->erase(stateSet_->iterator_to(*this));
435 lmm_variable_free(getModel()->getMaxminSystem(), getVariable());