1 /* Copyright (c) 2014-2018. The SimGrid Team. All rights reserved. */
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. */
6 #include "src/surf/network_ib.hpp"
7 #include "simgrid/sg_config.hpp"
8 #include "src/surf/HostImpl.hpp"
9 #include "src/surf/xml/platf.hpp"
10 #include "surf/surf.hpp"
12 #include <boost/algorithm/string/classification.hpp>
13 #include <boost/algorithm/string/split.hpp>
15 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(surf_network);
17 static void IB_create_host_callback(simgrid::s4u::Host& host){
18 using simgrid::kernel::resource::IBNode;
19 using simgrid::kernel::resource::NetworkIBModel;
22 // pour t->id -> rajouter une nouvelle struct dans le dict, pour stocker les comms actives
24 IBNode* act = new IBNode(id);
27 ((NetworkIBModel*)surf_network_model)->active_nodes.insert({host.get_name(), act});
30 static void IB_action_state_changed_callback(simgrid::kernel::resource::NetworkAction* action)
32 using simgrid::kernel::resource::IBNode;
33 using simgrid::kernel::resource::NetworkIBModel;
35 if (action->get_state() != simgrid::kernel::resource::Action::State::FINISHED)
37 std::pair<IBNode*,IBNode*> pair = ((NetworkIBModel*)surf_network_model)->active_comms[action];
38 XBT_DEBUG("IB callback - action %p finished", action);
40 ((NetworkIBModel*)surf_network_model)->updateIBfactors(action, pair.first, pair.second, 1);
42 ((NetworkIBModel*)surf_network_model)->active_comms.erase(action);
46 static void IB_action_init_callback(simgrid::kernel::resource::NetworkAction* action, simgrid::s4u::Host* src,
47 simgrid::s4u::Host* dst)
49 simgrid::kernel::resource::NetworkIBModel* ibModel = (simgrid::kernel::resource::NetworkIBModel*)surf_network_model;
50 simgrid::kernel::resource::IBNode* act_src;
51 simgrid::kernel::resource::IBNode* act_dst;
53 auto asrc = ibModel->active_nodes.find(src->get_name());
54 if (asrc != ibModel->active_nodes.end()) {
55 act_src = asrc->second;
57 throw std::out_of_range(std::string("Could not find '") + src->get_cname() + "' active comms !");
60 auto adst = ibModel->active_nodes.find(dst->get_name());
61 if (adst != ibModel->active_nodes.end()) {
62 act_dst = adst->second;
64 throw std::out_of_range(std::string("Could not find '") + dst->get_cname() + "' active comms !");
67 ibModel->active_comms[action]=std::make_pair(act_src, act_dst);
69 ibModel->updateIBfactors(action, act_src, act_dst, 0);
76 /************************************************************************/
77 /* New model based on MPI contention model for Infiniband platforms */
78 /************************************************************************/
79 /* @Inproceedings{mescal_vienne_phd, */
80 /* author={Jérôme Vienne}, */
81 /* title={prédiction de performances d’applications de calcul haute performance sur réseau Infiniband}, */
82 /* address={Grenoble FRANCE}, */
86 void surf_network_model_init_IB()
88 if (surf_network_model)
91 surf_network_model = new simgrid::kernel::resource::NetworkIBModel();
92 all_existing_models->push_back(surf_network_model);
93 simgrid::s4u::Link::on_communication_state_change.connect(IB_action_state_changed_callback);
94 simgrid::s4u::Link::on_communicate.connect(IB_action_init_callback);
95 simgrid::s4u::Host::on_creation.connect(IB_create_host_callback);
96 simgrid::config::set_default<double>("network/weight-S", 8775);
103 NetworkIBModel::NetworkIBModel() : NetworkSmpiModel()
105 std::string IB_factors_string = simgrid::config::get_value<std::string>("smpi/IB-penalty-factors");
106 std::vector<std::string> radical_elements;
107 boost::split(radical_elements, IB_factors_string, boost::is_any_of(";"));
109 surf_parse_assert(radical_elements.size() == 3, "smpi/IB-penalty-factors should be provided and contain 3 "
110 "elements, semi-colon separated. Example: 0.965;0.925;1.35");
113 Be = std::stod(radical_elements.front());
114 } catch (std::invalid_argument& ia) {
115 throw std::invalid_argument(std::string("First part of smpi/IB-penalty-factors is not numerical:") + ia.what());
119 Bs = std::stod(radical_elements.at(1));
120 } catch (std::invalid_argument& ia) {
121 throw std::invalid_argument(std::string("Second part of smpi/IB-penalty-factors is not numerical:") + ia.what());
125 ys = std::stod(radical_elements.back());
126 } catch (std::invalid_argument& ia) {
127 throw std::invalid_argument(std::string("Third part of smpi/IB-penalty-factors is not numerical:") + ia.what());
131 NetworkIBModel::~NetworkIBModel()
133 for (auto const& instance : active_nodes)
134 delete instance.second;
137 void NetworkIBModel::computeIBfactors(IBNode* root)
139 double num_comm_out = static_cast<double>(root->ActiveCommsUp.size());
140 double max_penalty_out = 0.0;
141 // first, compute all outbound penalties to get their max
142 for (std::vector<ActiveComm*>::iterator it = root->ActiveCommsUp.begin(); it != root->ActiveCommsUp.end(); ++it) {
143 double my_penalty_out = 1.0;
145 if (num_comm_out != 1) {
146 if ((*it)->destination->nbActiveCommsDown > 2) // number of comms sent to the receiving node
147 my_penalty_out = num_comm_out * Bs * ys;
149 my_penalty_out = num_comm_out * Bs;
152 max_penalty_out = std::max(max_penalty_out, my_penalty_out);
155 for (std::vector<ActiveComm*>::iterator it = root->ActiveCommsUp.begin(); it != root->ActiveCommsUp.end(); ++it) {
156 // compute inbound penalty
157 double my_penalty_in = 1.0;
158 int nb_comms = (*it)->destination->nbActiveCommsDown; // total number of incoming comms
160 my_penalty_in = ((*it)->destination->ActiveCommsDown)[root] // number of comm sent to dest by root node
161 * Be * (*it)->destination->ActiveCommsDown.size(); // number of different nodes sending to dest
163 double penalty = std::max(my_penalty_in, max_penalty_out);
165 double rate_before_update = (*it)->action->get_bound();
166 // save initial rate of the action
167 if ((*it)->init_rate == -1)
168 (*it)->init_rate = rate_before_update;
170 double penalized_bw = num_comm_out ? (*it)->init_rate / penalty : (*it)->init_rate;
172 if (not double_equals(penalized_bw, rate_before_update, sg_surf_precision)) {
173 XBT_DEBUG("%d->%d action %p penalty updated : bw now %f, before %f , initial rate %f", root->id,
174 (*it)->destination->id, (*it)->action, penalized_bw, (*it)->action->get_bound(), (*it)->init_rate);
175 get_maxmin_system()->update_variable_bound((*it)->action->get_variable(), penalized_bw);
177 XBT_DEBUG("%d->%d action %p penalty not updated : bw %f, initial rate %f", root->id, (*it)->destination->id,
178 (*it)->action, penalized_bw, (*it)->init_rate);
181 XBT_DEBUG("Finished computing IB penalties");
184 void NetworkIBModel::updateIBfactors_rec(IBNode* root, std::vector<bool>& updatedlist)
186 if (not updatedlist[root->id]) {
187 XBT_DEBUG("IB - Updating rec %d", root->id);
188 computeIBfactors(root);
189 updatedlist[root->id] = true;
190 for (std::vector<ActiveComm*>::iterator it = root->ActiveCommsUp.begin(); it != root->ActiveCommsUp.end(); ++it) {
191 if (not updatedlist[(*it)->destination->id])
192 updateIBfactors_rec((*it)->destination, updatedlist);
194 for (std::map<IBNode*, int>::iterator it = root->ActiveCommsDown.begin(); it != root->ActiveCommsDown.end(); ++it) {
195 if (not updatedlist[it->first->id])
196 updateIBfactors_rec(it->first, updatedlist);
201 void NetworkIBModel::updateIBfactors(NetworkAction* action, IBNode* from, IBNode* to, int remove)
203 if (from == to) // disregard local comms (should use loopback)
206 ActiveComm* comm = nullptr;
208 if (to->ActiveCommsDown[from] == 1)
209 to->ActiveCommsDown.erase(from);
211 to->ActiveCommsDown[from] -= 1;
213 to->nbActiveCommsDown--;
214 for (std::vector<ActiveComm*>::iterator it = from->ActiveCommsUp.begin(); it != from->ActiveCommsUp.end(); ++it) {
215 if ((*it)->action == action) {
217 from->ActiveCommsUp.erase(it);
224 ActiveComm* comm = new ActiveComm();
225 comm->action = action;
226 comm->destination = to;
227 from->ActiveCommsUp.push_back(comm);
229 to->ActiveCommsDown[from] += 1;
230 to->nbActiveCommsDown++;
232 XBT_DEBUG("IB - Updating %d", from->id);
233 std::vector<bool> updated(active_nodes.size(), false);
234 updateIBfactors_rec(from, updated);
235 XBT_DEBUG("IB - Finished updating %d", from->id);
240 } // namespace simgrid