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,
31 simgrid::kernel::resource::Action::State /*previous*/)
33 using simgrid::kernel::resource::IBNode;
34 using simgrid::kernel::resource::NetworkIBModel;
36 if (action->get_state() != simgrid::kernel::resource::Action::State::FINISHED)
38 std::pair<IBNode*,IBNode*> pair = ((NetworkIBModel*)surf_network_model)->active_comms[action];
39 XBT_DEBUG("IB callback - action %p finished", action);
41 ((NetworkIBModel*)surf_network_model)->updateIBfactors(action, pair.first, pair.second, 1);
43 ((NetworkIBModel*)surf_network_model)->active_comms.erase(action);
47 static void IB_action_init_callback(simgrid::kernel::resource::NetworkAction* action, simgrid::s4u::Host* src,
48 simgrid::s4u::Host* dst)
50 simgrid::kernel::resource::NetworkIBModel* ibModel = (simgrid::kernel::resource::NetworkIBModel*)surf_network_model;
51 simgrid::kernel::resource::IBNode* act_src;
52 simgrid::kernel::resource::IBNode* act_dst;
54 auto asrc = ibModel->active_nodes.find(src->get_name());
55 if (asrc != ibModel->active_nodes.end()) {
56 act_src = asrc->second;
58 throw std::out_of_range(std::string("Could not find '") + src->get_cname() + "' active comms !");
61 auto adst = ibModel->active_nodes.find(dst->get_name());
62 if (adst != ibModel->active_nodes.end()) {
63 act_dst = adst->second;
65 throw std::out_of_range(std::string("Could not find '") + dst->get_cname() + "' active comms !");
68 ibModel->active_comms[action]=std::make_pair(act_src, act_dst);
70 ibModel->updateIBfactors(action, act_src, act_dst, 0);
77 /************************************************************************/
78 /* New model based on MPI contention model for Infiniband platforms */
79 /************************************************************************/
80 /* @Inproceedings{mescal_vienne_phd, */
81 /* author={Jérôme Vienne}, */
82 /* title={prédiction de performances d’applications de calcul haute performance sur réseau Infiniband}, */
83 /* address={Grenoble FRANCE}, */
87 void surf_network_model_init_IB()
89 xbt_assert(surf_network_model == nullptr, "Cannot set the network model twice");
91 surf_network_model = new simgrid::kernel::resource::NetworkIBModel();
92 simgrid::s4u::Link::on_communication_state_change.connect(IB_action_state_changed_callback);
93 simgrid::s4u::Link::on_communicate.connect(IB_action_init_callback);
94 simgrid::s4u::Host::on_creation.connect(IB_create_host_callback);
95 simgrid::config::set_default<double>("network/weight-S", 8775);
102 NetworkIBModel::NetworkIBModel() : NetworkSmpiModel()
104 /* Do not add this into all_existing_models: our ancestor already does so */
106 std::string IB_factors_string = simgrid::config::get_value<std::string>("smpi/IB-penalty-factors");
107 std::vector<std::string> radical_elements;
108 boost::split(radical_elements, IB_factors_string, boost::is_any_of(";"));
110 surf_parse_assert(radical_elements.size() == 3, "smpi/IB-penalty-factors should be provided and contain 3 "
111 "elements, semi-colon separated. Example: 0.965;0.925;1.35");
114 Be = std::stod(radical_elements.front());
115 } catch (std::invalid_argument& ia) {
116 throw std::invalid_argument(std::string("First part of smpi/IB-penalty-factors is not numerical:") + ia.what());
120 Bs = std::stod(radical_elements.at(1));
121 } catch (std::invalid_argument& ia) {
122 throw std::invalid_argument(std::string("Second part of smpi/IB-penalty-factors is not numerical:") + ia.what());
126 ys = std::stod(radical_elements.back());
127 } catch (std::invalid_argument& ia) {
128 throw std::invalid_argument(std::string("Third part of smpi/IB-penalty-factors is not numerical:") + ia.what());
132 NetworkIBModel::~NetworkIBModel()
134 for (auto const& instance : active_nodes)
135 delete instance.second;
138 void NetworkIBModel::computeIBfactors(IBNode* root)
140 double num_comm_out = static_cast<double>(root->ActiveCommsUp.size());
141 double max_penalty_out = 0.0;
142 // first, compute all outbound penalties to get their max
143 for (std::vector<ActiveComm*>::iterator it = root->ActiveCommsUp.begin(); it != root->ActiveCommsUp.end(); ++it) {
144 double my_penalty_out = 1.0;
146 if (num_comm_out != 1) {
147 if ((*it)->destination->nbActiveCommsDown > 2) // number of comms sent to the receiving node
148 my_penalty_out = num_comm_out * Bs * ys;
150 my_penalty_out = num_comm_out * Bs;
153 max_penalty_out = std::max(max_penalty_out, my_penalty_out);
156 for (std::vector<ActiveComm*>::iterator it = root->ActiveCommsUp.begin(); it != root->ActiveCommsUp.end(); ++it) {
157 // compute inbound penalty
158 double my_penalty_in = 1.0;
159 int nb_comms = (*it)->destination->nbActiveCommsDown; // total number of incoming comms
161 my_penalty_in = ((*it)->destination->ActiveCommsDown)[root] // number of comm sent to dest by root node
162 * Be * (*it)->destination->ActiveCommsDown.size(); // number of different nodes sending to dest
164 double penalty = std::max(my_penalty_in, max_penalty_out);
166 double rate_before_update = (*it)->action->get_bound();
167 // save initial rate of the action
168 if ((*it)->init_rate == -1)
169 (*it)->init_rate = rate_before_update;
171 double penalized_bw = num_comm_out ? (*it)->init_rate / penalty : (*it)->init_rate;
173 if (not double_equals(penalized_bw, rate_before_update, sg_surf_precision)) {
174 XBT_DEBUG("%d->%d action %p penalty updated : bw now %f, before %f , initial rate %f", root->id,
175 (*it)->destination->id, (*it)->action, penalized_bw, (*it)->action->get_bound(), (*it)->init_rate);
176 get_maxmin_system()->update_variable_bound((*it)->action->get_variable(), penalized_bw);
178 XBT_DEBUG("%d->%d action %p penalty not updated : bw %f, initial rate %f", root->id, (*it)->destination->id,
179 (*it)->action, penalized_bw, (*it)->init_rate);
182 XBT_DEBUG("Finished computing IB penalties");
185 void NetworkIBModel::updateIBfactors_rec(IBNode* root, std::vector<bool>& updatedlist)
187 if (not updatedlist[root->id]) {
188 XBT_DEBUG("IB - Updating rec %d", root->id);
189 computeIBfactors(root);
190 updatedlist[root->id] = true;
191 for (std::vector<ActiveComm*>::iterator it = root->ActiveCommsUp.begin(); it != root->ActiveCommsUp.end(); ++it) {
192 if (not updatedlist[(*it)->destination->id])
193 updateIBfactors_rec((*it)->destination, updatedlist);
195 for (std::map<IBNode*, int>::iterator it = root->ActiveCommsDown.begin(); it != root->ActiveCommsDown.end(); ++it) {
196 if (not updatedlist[it->first->id])
197 updateIBfactors_rec(it->first, updatedlist);
202 void NetworkIBModel::updateIBfactors(NetworkAction* action, IBNode* from, IBNode* to, int remove)
204 if (from == to) // disregard local comms (should use loopback)
207 ActiveComm* comm = nullptr;
209 if (to->ActiveCommsDown[from] == 1)
210 to->ActiveCommsDown.erase(from);
212 to->ActiveCommsDown[from] -= 1;
214 to->nbActiveCommsDown--;
215 for (std::vector<ActiveComm*>::iterator it = from->ActiveCommsUp.begin(); it != from->ActiveCommsUp.end(); ++it) {
216 if ((*it)->action == action) {
218 from->ActiveCommsUp.erase(it);
225 ActiveComm* comm = new ActiveComm();
226 comm->action = action;
227 comm->destination = to;
228 from->ActiveCommsUp.push_back(comm);
230 to->ActiveCommsDown[from] += 1;
231 to->nbActiveCommsDown++;
233 XBT_DEBUG("IB - Updating %d", from->id);
234 std::vector<bool> updated(active_nodes.size(), false);
235 updateIBfactors_rec(from, updated);
236 XBT_DEBUG("IB - Finished updating %d", from->id);
241 } // namespace simgrid