1 /* Copyright (c) 2014-2023. 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 "src/kernel/EngineImpl.hpp"
9 #include "src/kernel/activity/CommImpl.hpp"
10 #include "src/kernel/resource/HostImpl.hpp"
11 #include "src/kernel/resource/models/network_ib.hpp"
12 #include "src/simgrid/math_utils.h"
13 #include "src/simgrid/module.hpp"
14 #include "src/simgrid/sg_config.hpp"
16 #include <boost/algorithm/string/classification.hpp>
17 #include <boost/algorithm/string/split.hpp>
19 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(res_network);
21 /****************************************************************/
22 /* Model based on MPI contention model for Infiniband platforms */
23 /****************************************************************/
24 /* @Inproceedings{mescal_vienne_phd, */
25 /* author={Jérôme Vienne}, */
26 /* title={prédiction de performances d’applications de calcul haute performance sur réseau Infiniband}, */
27 /* address={Grenoble FRANCE}, */
31 SIMGRID_REGISTER_NETWORK_MODEL(
33 "Realistic network model specifically tailored for HPC settings, with Infiniband contention model as described in "
34 "http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf",
36 using simgrid::kernel::resource::NetworkIBModel;
38 auto net_model = std::make_shared<NetworkIBModel>("Network_IB");
39 auto* engine = simgrid::kernel::EngineImpl::get_instance();
40 engine->add_model(net_model);
41 engine->get_netzone_root()->set_network_model(net_model);
43 simgrid::s4u::Link::on_communication_state_change_cb(NetworkIBModel::IB_action_state_changed_callback);
44 simgrid::s4u::Comm::on_start_cb(NetworkIBModel::IB_comm_start_callback);
45 simgrid::s4u::Host::on_creation_cb(NetworkIBModel::IB_create_host_callback);
46 simgrid::config::set_default<double>("network/weight-S", 8775);
49 namespace simgrid::kernel::resource {
51 void NetworkIBModel::IB_create_host_callback(s4u::Host const& host)
54 auto* ibModel = static_cast<NetworkIBModel*>(host.get_netpoint()->get_englobing_zone()->get_network_model().get());
55 ibModel->active_nodes.try_emplace(host.get_name(), id);
59 void NetworkIBModel::IB_action_state_changed_callback(NetworkAction& action, Action::State /*previous*/)
61 if (action.get_state() != Action::State::FINISHED)
63 auto* ibModel = static_cast<NetworkIBModel*>(action.get_model());
64 auto [src, dst] = ibModel->active_comms[&action];
66 XBT_DEBUG("IB callback - action %p finished", &action);
67 ibModel->update_IB_factors(&action, src, dst, true);
68 ibModel->active_comms.erase(&action);
71 void NetworkIBModel::IB_comm_start_callback(const s4u::Comm& comm)
73 auto* action = static_cast<NetworkAction*>(static_cast<activity::CommImpl*>(comm.get_impl())->model_action_);
74 auto* ibModel = static_cast<NetworkIBModel*>(action->get_model());
75 auto* act_src = &ibModel->active_nodes.at(action->get_src().get_name());
76 auto* act_dst = &ibModel->active_nodes.at(action->get_dst().get_name());
78 ibModel->active_comms[action] = std::make_pair(act_src, act_dst);
79 ibModel->update_IB_factors(action, act_src, act_dst, false);
82 NetworkIBModel::NetworkIBModel(const std::string& name) : NetworkCm02Model(name)
84 std::string IB_factors_string = config::get_value<std::string>("smpi/IB-penalty-factors");
85 std::vector<std::string> radical_elements;
86 boost::split(radical_elements, IB_factors_string, boost::is_any_of(";"));
88 xbt_assert(radical_elements.size() == 3, "smpi/IB-penalty-factors should be provided and contain 3 "
89 "elements, semi-colon separated. Example: 0.965;0.925;1.35");
92 Be_ = std::stod(radical_elements.front());
93 } catch (const std::invalid_argument& ia) {
94 throw std::invalid_argument(std::string("First part of smpi/IB-penalty-factors is not numerical:") + ia.what());
98 Bs_ = std::stod(radical_elements.at(1));
99 } catch (const std::invalid_argument& ia) {
100 throw std::invalid_argument(std::string("Second part of smpi/IB-penalty-factors is not numerical:") + ia.what());
104 ys_ = std::stod(radical_elements.back());
105 } catch (const std::invalid_argument& ia) {
106 throw std::invalid_argument(std::string("Third part of smpi/IB-penalty-factors is not numerical:") + ia.what());
110 void NetworkIBModel::compute_IB_factors(IBNode* root) const
112 size_t num_comm_out = root->active_comms_up_.size();
113 double max_penalty_out = 0.0;
114 // first, compute all outbound penalties to get their max
115 for (ActiveComm const* comm : root->active_comms_up_) {
116 double my_penalty_out = 1.0;
118 if (num_comm_out != 1) {
119 if (comm->destination->nb_active_comms_down_ > 2) // number of comms sent to the receiving node
120 my_penalty_out = num_comm_out * Bs_ * ys_;
122 my_penalty_out = num_comm_out * Bs_;
125 max_penalty_out = std::max(max_penalty_out, my_penalty_out);
128 for (ActiveComm* comm : root->active_comms_up_) {
129 // compute inbound penalty
130 double my_penalty_in = 1.0;
131 if (comm->destination->nb_active_comms_down_ != 1) // total number of incoming comms
132 my_penalty_in = (comm->destination->active_comms_down_)[root] // number of comm sent to dest by root node
133 * Be_ * comm->destination->active_comms_down_.size(); // number of different nodes sending to dest
135 double penalty = std::max(my_penalty_in, max_penalty_out);
137 double rate_before_update = comm->action->get_bound();
138 // save initial rate of the action
139 if (comm->init_rate == -1)
140 comm->init_rate = rate_before_update;
142 double penalized_bw = num_comm_out ? comm->init_rate / penalty : comm->init_rate;
144 if (not double_equals(penalized_bw, rate_before_update, sg_precision_timing)) {
145 XBT_DEBUG("%d->%d action %p penalty updated : bw now %f, before %f , initial rate %f", root->id_,
146 comm->destination->id_, comm->action, penalized_bw, comm->action->get_bound(), comm->init_rate);
147 get_maxmin_system()->update_variable_bound(comm->action->get_variable(), penalized_bw);
149 XBT_DEBUG("%d->%d action %p penalty not updated : bw %f, initial rate %f", root->id_, comm->destination->id_,
150 comm->action, penalized_bw, comm->init_rate);
153 XBT_DEBUG("Finished computing IB penalties");
156 void NetworkIBModel::update_IB_factors_rec(IBNode* root, std::vector<bool>& updatedlist) const
158 if (not updatedlist[root->id_]) {
159 XBT_DEBUG("IB - Updating rec %d", root->id_);
160 compute_IB_factors(root);
161 updatedlist[root->id_] = true;
162 for (ActiveComm const* comm : root->active_comms_up_) {
163 if (not updatedlist[comm->destination->id_])
164 update_IB_factors_rec(comm->destination, updatedlist);
166 for (auto const& [comm, _] : root->active_comms_down_) {
167 if (not updatedlist[comm->id_])
168 update_IB_factors_rec(comm, updatedlist);
173 void NetworkIBModel::update_IB_factors(NetworkAction* action, IBNode* from, IBNode* to, bool remove) const
175 if (from == to) // disregard local comms (should use loopback)
179 if (to->active_comms_down_[from] == 1)
180 to->active_comms_down_.erase(from);
182 to->active_comms_down_[from] -= 1;
184 to->nb_active_comms_down_--;
185 if (auto it = std::find_if(begin(from->active_comms_up_), end(from->active_comms_up_),
186 [action](const ActiveComm* comm) { return comm->action == action; });
187 it != std::end(from->active_comms_up_)) {
189 from->active_comms_up_.erase(it);
194 auto* comm = new ActiveComm();
195 comm->action = action;
196 comm->destination = to;
197 from->active_comms_up_.push_back(comm);
199 to->active_comms_down_[from] += 1;
200 to->nb_active_comms_down_++;
202 XBT_DEBUG("IB - Updating %d", from->id_);
203 std::vector<bool> updated(active_nodes.size(), false);
204 update_IB_factors_rec(from, updated);
205 XBT_DEBUG("IB - Finished updating %d", from->id_);
207 } // namespace simgrid::kernel::resource