1 /* Copyright (c) 2014-2022. 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 <simgrid/kernel/routing/NetPoint.hpp>
8 #include "simgrid/sg_config.hpp"
9 #include "src/kernel/EngineImpl.hpp"
10 #include "src/kernel/activity/CommImpl.hpp"
11 #include "src/surf/HostImpl.hpp"
12 #include "src/surf/network_ib.hpp"
14 #include <boost/algorithm/string/classification.hpp>
15 #include <boost/algorithm/string/split.hpp>
17 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(res_network);
23 /************************************************************************/
24 /* New model based on MPI contention model for Infiniband platforms */
25 /************************************************************************/
26 /* @Inproceedings{mescal_vienne_phd, */
27 /* author={Jérôme Vienne}, */
28 /* title={prédiction de performances d’applications de calcul haute performance sur réseau Infiniband}, */
29 /* address={Grenoble FRANCE}, */
33 void surf_network_model_init_IB()
35 using simgrid::kernel::resource::NetworkIBModel;
37 auto net_model = std::make_shared<NetworkIBModel>("Network_IB");
38 auto* engine = simgrid::kernel::EngineImpl::get_instance();
39 engine->add_model(net_model);
40 engine->get_netzone_root()->set_network_model(net_model);
42 simgrid::s4u::Link::on_communication_state_change_cb(NetworkIBModel::IB_action_state_changed_callback);
43 simgrid::kernel::activity::CommImpl::on_start.connect(NetworkIBModel::IB_comm_start_callback);
44 simgrid::s4u::Host::on_creation_cb(NetworkIBModel::IB_create_host_callback);
45 simgrid::config::set_default<double>("network/weight-S", 8775);
52 void NetworkIBModel::IB_create_host_callback(s4u::Host const& host)
55 auto* ibModel = static_cast<NetworkIBModel*>(host.get_netpoint()->get_englobing_zone()->get_network_model().get());
56 ibModel->active_nodes.emplace(host.get_name(), IBNode(id));
60 void NetworkIBModel::IB_action_state_changed_callback(NetworkAction& action, Action::State /*previous*/)
62 if (action.get_state() != Action::State::FINISHED)
64 auto* ibModel = static_cast<NetworkIBModel*>(action.get_model());
65 std::pair<IBNode*, IBNode*> pair = ibModel->active_comms[&action];
67 XBT_DEBUG("IB callback - action %p finished", &action);
68 ibModel->update_IB_factors(&action, pair.first, pair.second, 1);
69 ibModel->active_comms.erase(&action);
72 void NetworkIBModel::IB_comm_start_callback(const activity::CommImpl& comm)
74 auto* action = static_cast<NetworkAction*>(comm.surf_action_);
75 auto* ibModel = static_cast<NetworkIBModel*>(action->get_model());
76 auto* act_src = &ibModel->active_nodes.at(action->get_src().get_name());
77 auto* act_dst = &ibModel->active_nodes.at(action->get_dst().get_name());
79 ibModel->active_comms[action] = std::make_pair(act_src, act_dst);
80 ibModel->update_IB_factors(action, act_src, act_dst, 0);
83 NetworkIBModel::NetworkIBModel(const std::string& name) : NetworkSmpiModel(name)
85 std::string IB_factors_string = config::get_value<std::string>("smpi/IB-penalty-factors");
86 std::vector<std::string> radical_elements;
87 boost::split(radical_elements, IB_factors_string, boost::is_any_of(";"));
89 xbt_assert(radical_elements.size() == 3, "smpi/IB-penalty-factors should be provided and contain 3 "
90 "elements, semi-colon separated. Example: 0.965;0.925;1.35");
93 Be_ = std::stod(radical_elements.front());
94 } catch (const std::invalid_argument& ia) {
95 throw std::invalid_argument(std::string("First part of smpi/IB-penalty-factors is not numerical:") + ia.what());
99 Bs_ = std::stod(radical_elements.at(1));
100 } catch (const std::invalid_argument& ia) {
101 throw std::invalid_argument(std::string("Second part of smpi/IB-penalty-factors is not numerical:") + ia.what());
105 ys_ = std::stod(radical_elements.back());
106 } catch (const std::invalid_argument& ia) {
107 throw std::invalid_argument(std::string("Third part of smpi/IB-penalty-factors is not numerical:") + ia.what());
111 void NetworkIBModel::compute_IB_factors(IBNode* root) const
113 size_t num_comm_out = root->active_comms_up_.size();
114 double max_penalty_out = 0.0;
115 // first, compute all outbound penalties to get their max
116 for (ActiveComm const* comm : root->active_comms_up_) {
117 double my_penalty_out = 1.0;
119 if (num_comm_out != 1) {
120 if (comm->destination->nb_active_comms_down_ > 2) // number of comms sent to the receiving node
121 my_penalty_out = num_comm_out * Bs_ * ys_;
123 my_penalty_out = num_comm_out * Bs_;
126 max_penalty_out = std::max(max_penalty_out, my_penalty_out);
129 for (ActiveComm* comm : root->active_comms_up_) {
130 // compute inbound penalty
131 double my_penalty_in = 1.0;
132 int nb_comms = comm->destination->nb_active_comms_down_; // total number of incoming comms
134 my_penalty_in = (comm->destination->active_comms_down_)[root] // number of comm sent to dest by root node
135 * Be_ * comm->destination->active_comms_down_.size(); // number of different nodes sending to dest
137 double penalty = std::max(my_penalty_in, max_penalty_out);
139 double rate_before_update = comm->action->get_bound();
140 // save initial rate of the action
141 if (comm->init_rate == -1)
142 comm->init_rate = rate_before_update;
144 double penalized_bw = num_comm_out ? comm->init_rate / penalty : comm->init_rate;
146 if (not double_equals(penalized_bw, rate_before_update, sg_surf_precision)) {
147 XBT_DEBUG("%d->%d action %p penalty updated : bw now %f, before %f , initial rate %f", root->id_,
148 comm->destination->id_, comm->action, penalized_bw, comm->action->get_bound(), comm->init_rate);
149 get_maxmin_system()->update_variable_bound(comm->action->get_variable(), penalized_bw);
151 XBT_DEBUG("%d->%d action %p penalty not updated : bw %f, initial rate %f", root->id_, comm->destination->id_,
152 comm->action, penalized_bw, comm->init_rate);
155 XBT_DEBUG("Finished computing IB penalties");
158 void NetworkIBModel::update_IB_factors_rec(IBNode* root, std::vector<bool>& updatedlist) const
160 if (not updatedlist[root->id_]) {
161 XBT_DEBUG("IB - Updating rec %d", root->id_);
162 compute_IB_factors(root);
163 updatedlist[root->id_] = true;
164 for (ActiveComm const* comm : root->active_comms_up_) {
165 if (not updatedlist[comm->destination->id_])
166 update_IB_factors_rec(comm->destination, updatedlist);
168 for (std::map<IBNode*, int>::value_type const& comm : root->active_comms_down_) {
169 if (not updatedlist[comm.first->id_])
170 update_IB_factors_rec(comm.first, updatedlist);
175 void NetworkIBModel::update_IB_factors(NetworkAction* action, IBNode* from, IBNode* to, int remove) const
177 if (from == to) // disregard local comms (should use loopback)
181 if (to->active_comms_down_[from] == 1)
182 to->active_comms_down_.erase(from);
184 to->active_comms_down_[from] -= 1;
186 to->nb_active_comms_down_--;
187 auto it = std::find_if(begin(from->active_comms_up_), end(from->active_comms_up_),
188 [action](const ActiveComm* comm) { return comm->action == action; });
189 if (it != std::end(from->active_comms_up_)) {
191 from->active_comms_up_.erase(it);
196 auto* comm = new ActiveComm();
197 comm->action = action;
198 comm->destination = to;
199 from->active_comms_up_.push_back(comm);
201 to->active_comms_down_[from] += 1;
202 to->nb_active_comms_down_++;
204 XBT_DEBUG("IB - Updating %d", from->id_);
205 std::vector<bool> updated(active_nodes.size(), false);
206 update_IB_factors_rec(from, updated);
207 XBT_DEBUG("IB - Finished updating %d", from->id_);
209 } // namespace resource
210 } // namespace kernel
211 } // namespace simgrid