1 /* Copyright (c) 2014-2015. 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 "network_ib.hpp"
11 #include "simgrid/sg_config.h"
12 #include "maxmin_private.hpp"
13 #include "src/surf/host_interface.hpp"
15 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(surf_network);
17 static void IB_create_host_callback(Host* host){
20 // pour t->id -> rajouter une nouvelle struct dans le dict, pour stocker les comms actives
21 if(((NetworkIBModel*)surf_network_model)->active_nodes==NULL)
22 ((NetworkIBModel*)surf_network_model)->active_nodes=xbt_dict_new();
24 IBNode* act = new IBNode(id);
27 xbt_dict_set(((NetworkIBModel*)surf_network_model)->active_nodes,
28 host->getName(), act, NULL);
32 static void IB_action_state_changed_callback(NetworkAction *action, e_surf_action_state_t statein, e_surf_action_state_t stateout){
33 if(statein!=SURF_ACTION_RUNNING|| stateout!=SURF_ACTION_DONE)
35 std::pair<IBNode*,IBNode*> pair = ((NetworkIBModel*)surf_network_model)->active_comms[action];
36 XBT_DEBUG("IB callback - action %p finished", action);
38 ((NetworkIBModel*)surf_network_model)->updateIBfactors(action, pair.first, pair.second, 1);
40 ((NetworkIBModel*)surf_network_model)->active_comms.erase(action);
45 static void IB_action_init_callback(NetworkAction *action,RoutingEdge *src, RoutingEdge *dst, double size, double rate){
46 if(((NetworkIBModel*)surf_network_model)->active_nodes==NULL)
47 xbt_die("IB comm added, without any node connected !");
49 IBNode* act_src= (IBNode*) xbt_dict_get_or_null(((NetworkIBModel*)surf_network_model)->active_nodes, src->getName());
51 xbt_die("could not find src node active comms !");
54 IBNode* act_dst= (IBNode*) xbt_dict_get_or_null(((NetworkIBModel*)surf_network_model)->active_nodes, dst->getName());
56 xbt_die("could not find dst node active comms !");
57 // act_dst->rate=rate;
59 ((NetworkIBModel*)surf_network_model)->active_comms[action]=std::make_pair(act_src, act_dst);
60 //post the action in the second dist, to retrieve in the other callback
61 XBT_DEBUG("IB callback - action %p init", action);
63 ((NetworkIBModel*)surf_network_model)->updateIBfactors(action, act_src, act_dst, 0);
73 /************************************************************************/
74 /* New model based on MPI contention model for Infiniband platforms */
75 /************************************************************************/
76 /* @Inproceedings{mescal_vienne_phd, */
77 /* author={Jérôme Vienne}, */
78 /* title={prédiction de performances d’applications de calcul haute performance sur réseau Infiniband}, */
79 /* address={Grenoble FRANCE}, */
83 void surf_network_model_init_IB(void)
86 if (surf_network_model)
88 surf_network_model = new NetworkIBModel();
89 net_define_callbacks();
90 xbt_dynar_push(all_existing_models, &surf_network_model);
91 surf_callback_connect(networkActionStateChangedCallbacks, IB_action_state_changed_callback);
92 surf_callback_connect(networkCommunicateCallbacks, IB_action_init_callback);
94 hostCreatedCallbacks.connect(IB_create_host_callback);
95 xbt_cfg_setdefault_double(_sg_cfg_set, "network/weight_S", 8775);
99 NetworkIBModel::NetworkIBModel()
100 : NetworkSmpiModel() {
104 const char* IB_factors_string=sg_cfg_get_string("smpi/IB_penalty_factors");
105 xbt_dynar_t radical_elements = xbt_str_split(IB_factors_string, ";");
107 if(xbt_dynar_length(radical_elements)!=3)
108 surf_parse_error("smpi/IB_penalty_factors should be provided and contain 3 elements, semi-colon separated : for example 0.965;0.925;1.35");
110 Be = atof(xbt_dynar_get_as(radical_elements, 0, char *));
111 Bs = atof(xbt_dynar_get_as(radical_elements, 1, char *));
112 ys = atof(xbt_dynar_get_as(radical_elements, 2, char *));
114 xbt_dynar_free(&radical_elements);
117 NetworkIBModel::~NetworkIBModel()
119 xbt_dict_cursor_t cursor = NULL;
120 IBNode* instance = NULL;
122 xbt_dict_foreach(active_nodes, cursor, name, instance)
124 xbt_dict_free(&active_nodes);
127 void NetworkIBModel::computeIBfactors(IBNode *root) {
128 double penalized_bw=0.0;
129 double num_comm_out = (double) root->ActiveCommsUp.size();
130 double max_penalty_out=0.0;
131 //first, compute all outbound penalties to get their max
132 for (std::vector<ActiveComm*>::iterator it= root->ActiveCommsUp.begin(); it != root->ActiveCommsUp.end(); ++it) {
133 double my_penalty_out = 1.0;
136 if((*it)->destination->nbActiveCommsDown > 2)//number of comms sent to the receiving node
137 my_penalty_out = num_comm_out * Bs * ys;
139 my_penalty_out = num_comm_out * Bs;
142 max_penalty_out = std::max(max_penalty_out,my_penalty_out);
145 for (std::vector<ActiveComm*>::iterator it= root->ActiveCommsUp.begin(); it != root->ActiveCommsUp.end(); ++it) {
147 //compute inbound penalty
148 double my_penalty_in = 1.0;
149 int nb_comms = (*it)->destination->nbActiveCommsDown;//total number of incoming comms
151 my_penalty_in = ((*it)->destination->ActiveCommsDown)[root] //number of comm sent to dest by root node
153 * (*it)->destination->ActiveCommsDown.size();//number of different nodes sending to dest
155 double penalty = std::max(my_penalty_in,max_penalty_out);
157 double rate_before_update = (*it)->action->getBound();
158 //save initial rate of the action
159 if((*it)->init_rate==-1)
160 (*it)->init_rate= rate_before_update;
162 penalized_bw= ! num_comm_out ? (*it)->init_rate : (*it)->init_rate /penalty;
164 if (!double_equals(penalized_bw, rate_before_update, sg_surf_precision)){
165 XBT_DEBUG("%d->%d action %p penalty updated : bw now %f, before %f , initial rate %f", root->id,(*it)->destination->id,(*it)->action,penalized_bw, (*it)->action->getBound(), (*it)->init_rate );
166 lmm_update_variable_bound(p_maxminSystem, (*it)->action->getVariable(), penalized_bw);
168 XBT_DEBUG("%d->%d action %p penalty not updated : bw %f, initial rate %f", root->id,(*it)->destination->id,(*it)->action,penalized_bw, (*it)->init_rate );
172 XBT_DEBUG("Finished computing IB penalties");
175 void NetworkIBModel::updateIBfactors_rec(IBNode *root, bool* updatedlist) {
176 if(updatedlist[root->id]==0){
177 XBT_DEBUG("IB - Updating rec %d", root->id);
178 computeIBfactors(root);
179 updatedlist[root->id]=1;
180 for (std::vector<ActiveComm*>::iterator it= root->ActiveCommsUp.begin(); it != root->ActiveCommsUp.end(); ++it) {
181 if(updatedlist[(*it)->destination->id]!=1)
182 updateIBfactors_rec((*it)->destination, updatedlist);
184 for (std::map<IBNode*, int>::iterator it= root->ActiveCommsDown.begin(); it != root->ActiveCommsDown.end(); ++it) {
185 if(updatedlist[it->first->id]!=1)
186 updateIBfactors_rec(it->first, updatedlist);
192 void NetworkIBModel::updateIBfactors(NetworkAction *action, IBNode *from, IBNode * to, int remove) {
193 if (from == to)//disregard local comms (should use loopback)
196 bool* updated=(bool*)xbt_malloc0(xbt_dict_size(active_nodes)*sizeof(bool));
197 ActiveComm* comm=NULL;
199 if(to->ActiveCommsDown[from]==1)
200 to->ActiveCommsDown.erase(from);
202 to->ActiveCommsDown[from]-=1;
204 to->nbActiveCommsDown--;
205 for (std::vector<ActiveComm*>::iterator it= from->ActiveCommsUp.begin();
206 it != from->ActiveCommsUp.end(); ++it) {
207 if((*it)->action==action){
209 from->ActiveCommsUp.erase(it);
217 ActiveComm* comm=new ActiveComm();
219 comm->destination=to;
220 from->ActiveCommsUp.push_back(comm);
222 to->ActiveCommsDown[from]+=1;
223 to->nbActiveCommsDown++;
225 XBT_DEBUG("IB - Updating %d", from->id);
226 updateIBfactors_rec(from, updated);
227 XBT_DEBUG("IB - Finished updating %d", from->id);