1 /* Copyright (c) 2014-2016. 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/kernel/routing/DragonflyZone.hpp"
7 #include "src/kernel/routing/NetPoint.hpp"
8 #include "src/surf/network_interface.hpp"
10 #include <boost/algorithm/string/classification.hpp>
11 #include <boost/algorithm/string/split.hpp>
14 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_route_cluster_dragonfly, surf_route_cluster, "Dragonfly Routing part of surf");
20 DragonflyZone::DragonflyZone(NetZone* father, const char* name) : ClusterZone(father, name)
24 DragonflyZone::~DragonflyZone()
26 if (this->routers_ != nullptr) {
27 for (unsigned int i = 0; i < this->numGroups_ * this->numChassisPerGroup_ * this->numBladesPerChassis_; i++)
33 void DragonflyZone::rankId_to_coords(int rankId, unsigned int (*coords)[4])
35 // coords : group, chassis, blade, node
36 (*coords)[0] = rankId / (numChassisPerGroup_ * numBladesPerChassis_ * numNodesPerBlade_);
37 rankId = rankId % (numChassisPerGroup_ * numBladesPerChassis_ * numNodesPerBlade_);
38 (*coords)[1] = rankId / (numBladesPerChassis_ * numNodesPerBlade_);
39 rankId = rankId % (numBladesPerChassis_ * numNodesPerBlade_);
40 (*coords)[2] = rankId / numNodesPerBlade_;
41 (*coords)[3] = rankId % numNodesPerBlade_;
44 void DragonflyZone::parse_specific_arguments(sg_platf_cluster_cbarg_t cluster)
46 std::vector<std::string> parameters;
47 std::vector<std::string> tmp;
48 boost::split(parameters, cluster->topo_parameters, boost::is_any_of(";"));
50 // TODO : we have to check for zeros and negative numbers, or it might crash
51 if (parameters.size() != 4) {
53 "Dragonfly are defined by the number of groups, chassis per groups, blades per chassis, nodes per blade");
56 // Blue network : number of groups, number of links between each group
57 boost::split(tmp, parameters[0], boost::is_any_of(","));
58 if (tmp.size() != 2) {
59 surf_parse_error("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
62 this->numGroups_ = xbt_str_parse_int(tmp[0].c_str(), "Invalid number of groups: %s");
63 this->numLinksBlue_ = xbt_str_parse_int(tmp[1].c_str(), "Invalid number of links for the blue level: %s");
65 // Black network : number of chassis/group, number of links between each router on the black network
66 boost::split(tmp, parameters[1], boost::is_any_of(","));
67 if (tmp.size() != 2) {
68 surf_parse_error("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
71 this->numChassisPerGroup_ = xbt_str_parse_int(tmp[0].c_str(), "Invalid number of groups: %s");
72 this->numLinksBlack_ = xbt_str_parse_int(tmp[1].c_str(), "Invalid number of links for the black level: %s");
74 // Green network : number of blades/chassis, number of links between each router on the green network
75 boost::split(tmp, parameters[2], boost::is_any_of(","));
76 if (tmp.size() != 2) {
77 surf_parse_error("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
80 this->numBladesPerChassis_ = xbt_str_parse_int(tmp[0].c_str(), "Invalid number of groups: %s");
81 this->numLinksGreen_ = xbt_str_parse_int(tmp[1].c_str(), "Invalid number of links for the green level: %s");
83 // The last part of topo_parameters should be the number of nodes per blade
84 this->numNodesPerBlade_ =
85 xbt_str_parse_int(parameters[3].c_str(), "Last parameter is not the amount of nodes per blade: %s");
86 this->cluster_ = cluster;
90 * Generate the cluster once every node is created
92 void DragonflyZone::seal()
94 if (this->numNodesPerBlade_ == 0) {
98 this->generateRouters();
99 this->generateLinks();
102 DragonflyRouter::DragonflyRouter(int group, int chassis, int blade) : group_(group), chassis_(chassis), blade_(blade)
106 DragonflyRouter::~DragonflyRouter()
108 if (this->myNodes_ != nullptr)
110 if (this->greenLinks_ != nullptr)
111 xbt_free(greenLinks_);
112 if (this->blackLinks_ != nullptr)
113 xbt_free(blackLinks_);
114 if (this->blueLinks_ != nullptr)
115 xbt_free(blueLinks_);
118 void DragonflyZone::generateRouters()
120 this->routers_ = static_cast<DragonflyRouter**>(xbt_malloc0(this->numGroups_ * this->numChassisPerGroup_ *
121 this->numBladesPerChassis_ * sizeof(DragonflyRouter*)));
123 for (unsigned int i = 0; i < this->numGroups_; i++) {
124 for (unsigned int j = 0; j < this->numChassisPerGroup_; j++) {
125 for (unsigned int k = 0; k < this->numBladesPerChassis_; k++) {
126 DragonflyRouter* router = new DragonflyRouter(i, j, k);
127 this->routers_[i * this->numChassisPerGroup_ * this->numBladesPerChassis_ + j * this->numBladesPerChassis_ +
134 void DragonflyZone::createLink(std::string id, int numlinks, surf::LinkImpl** linkup, surf::LinkImpl** linkdown)
138 LinkCreationArgs linkTemplate;
139 linkTemplate.bandwidth = this->cluster_->bw * numlinks;
140 linkTemplate.latency = this->cluster_->lat;
141 linkTemplate.policy = this->cluster_->sharing_policy; // sthg to do with that ?
142 linkTemplate.id = id;
143 sg_platf_new_link(&linkTemplate);
144 XBT_DEBUG("Generating link %s", id.c_str());
145 surf::LinkImpl* link;
147 if (this->cluster_->sharing_policy == SURF_LINK_FULLDUPLEX) {
148 tmpID = linkTemplate.id + "_UP";
149 link = surf::LinkImpl::byName(tmpID.c_str());
150 *linkup = link; // check link?
151 tmpID = linkTemplate.id + "_DOWN";
152 link = surf::LinkImpl::byName(tmpID.c_str());
153 *linkdown = link; // check link ?
155 link = surf::LinkImpl::byName(linkTemplate.id.c_str());
161 void DragonflyZone::generateLinks()
163 static int uniqueId = 0;
164 surf::LinkImpl* linkup;
165 surf::LinkImpl* linkdown;
167 unsigned int numRouters = this->numGroups_ * this->numChassisPerGroup_ * this->numBladesPerChassis_;
169 if (this->cluster_->sharing_policy == SURF_LINK_FULLDUPLEX)
170 numLinksperLink_ = 2;
172 // Links from routers to their local nodes.
173 for (unsigned int i = 0; i < numRouters; i++) {
174 // allocate structures
175 this->routers_[i]->myNodes_ = static_cast<surf::LinkImpl**>(
176 xbt_malloc0(numLinksperLink_ * this->numNodesPerBlade_ * sizeof(surf::LinkImpl*)));
177 this->routers_[i]->greenLinks_ =
178 static_cast<surf::LinkImpl**>(xbt_malloc0(this->numBladesPerChassis_ * sizeof(surf::LinkImpl*)));
179 this->routers_[i]->blackLinks_ =
180 static_cast<surf::LinkImpl**>(xbt_malloc0(this->numChassisPerGroup_ * sizeof(surf::LinkImpl*)));
182 for (unsigned int j = 0; j < numLinksperLink_ * this->numNodesPerBlade_; j += numLinksperLink_) {
183 std::string id = "local_link_from_router_"+ std::to_string(i) + "_to_node_" +
184 std::to_string(j / numLinksperLink_) + "_" + std::to_string(uniqueId);
185 this->createLink(id, 1, &linkup, &linkdown);
187 if (this->cluster_->sharing_policy == SURF_LINK_FULLDUPLEX) {
188 this->routers_[i]->myNodes_[j] = linkup;
189 this->routers_[i]->myNodes_[j + 1] = linkdown;
191 this->routers_[i]->myNodes_[j] = linkup;
197 // Green links from routers to same chassis routers - alltoall
198 for (unsigned int i = 0; i < this->numGroups_ * this->numChassisPerGroup_; i++) {
199 for (unsigned int j = 0; j < this->numBladesPerChassis_; j++) {
200 for (unsigned int k = j + 1; k < this->numBladesPerChassis_; k++) {
201 std::string id = "green_link_in_chassis_" + std::to_string(i % numChassisPerGroup_) +"_between_routers_" +
202 std::to_string(j) + "_and_" + std::to_string(k) + "_" + std::to_string(uniqueId);
203 this->createLink(id, this->numLinksGreen_, &linkup, &linkdown);
205 this->routers_[i * numBladesPerChassis_ + j]->greenLinks_[k] = linkup;
206 this->routers_[i * numBladesPerChassis_ + k]->greenLinks_[j] = linkdown;
212 // Black links from routers to same group routers - alltoall
213 for (unsigned int i = 0; i < this->numGroups_; i++) {
214 for (unsigned int j = 0; j < this->numChassisPerGroup_; j++) {
215 for (unsigned int k = j + 1; k < this->numChassisPerGroup_; k++) {
216 for (unsigned int l = 0; l < this->numBladesPerChassis_; l++) {
217 std::string id = "black_link_in_group_" + std::to_string(i) + "_between_chassis_" + std::to_string(j) +
218 "_and_" + std::to_string(k) +"_blade_" + std::to_string(l) + "_" + std::to_string(uniqueId);
219 this->createLink(id, this->numLinksBlack_, &linkup, &linkdown);
221 this->routers_[i * numBladesPerChassis_ * numChassisPerGroup_ + j * numBladesPerChassis_ + l]
222 ->blackLinks_[k] = linkup;
223 this->routers_[i * numBladesPerChassis_ * numChassisPerGroup_ + k * numBladesPerChassis_ + l]
224 ->blackLinks_[j] = linkdown;
231 // Blue links between groups - Not all routers involved, only one per group is linked to others. Let's say router n of
232 // each group is linked to group n.
233 // FIXME: in reality blue links may be attached to several different routers
234 for (unsigned int i = 0; i < this->numGroups_; i++) {
235 for (unsigned int j = i + 1; j < this->numGroups_; j++) {
236 unsigned int routernumi = i * numBladesPerChassis_ * numChassisPerGroup_ + j;
237 unsigned int routernumj = j * numBladesPerChassis_ * numChassisPerGroup_ + i;
238 this->routers_[routernumi]->blueLinks_ = static_cast<surf::LinkImpl**>(xbt_malloc0(sizeof(surf::LinkImpl*)));
239 this->routers_[routernumj]->blueLinks_ = static_cast<surf::LinkImpl**>(xbt_malloc0(sizeof(surf::LinkImpl*)));
240 std::string id = "blue_link_between_group_"+ std::to_string(i) +"_and_" + std::to_string(j) +"_routers_" +
241 std::to_string(routernumi) + "_and_" + std::to_string(routernumj) + "_" + std::to_string(uniqueId);
242 this->createLink(id, this->numLinksBlue_, &linkup, &linkdown);
244 this->routers_[routernumi]->blueLinks_[0] = linkup;
245 this->routers_[routernumj]->blueLinks_[0] = linkdown;
251 void DragonflyZone::getLocalRoute(NetPoint* src, NetPoint* dst, sg_platf_route_cbarg_t route, double* latency)
253 // Minimal routing version.
254 // TODO : non-minimal random one, and adaptive ?
256 if (dst->isRouter() || src->isRouter())
259 XBT_VERB("dragonfly getLocalRout from '%s'[%d] to '%s'[%d]", src->name().c_str(), src->id(), dst->name().c_str(),
262 if ((src->id() == dst->id()) && hasLoopback_) {
263 std::pair<surf::LinkImpl*, surf::LinkImpl*> info = privateLinks_.at(src->id() * linkCountPerNode_);
265 route->link_list->push_back(info.first);
267 *latency += info.first->latency();
271 unsigned int myCoords[4];
272 rankId_to_coords(src->id(), &myCoords);
273 unsigned int targetCoords[4];
274 rankId_to_coords(dst->id(), &targetCoords);
275 XBT_DEBUG("src : %u group, %u chassis, %u blade, %u node", myCoords[0], myCoords[1], myCoords[2], myCoords[3]);
276 XBT_DEBUG("dst : %u group, %u chassis, %u blade, %u node", targetCoords[0], targetCoords[1], targetCoords[2],
279 DragonflyRouter* myRouter = routers_[myCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
280 myCoords[1] * numBladesPerChassis_ + myCoords[2]];
281 DragonflyRouter* targetRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
282 targetCoords[1] * numBladesPerChassis_ + targetCoords[2]];
283 DragonflyRouter* currentRouter = myRouter;
285 // node->router local link
286 route->link_list->push_back(myRouter->myNodes_[myCoords[3] * numLinksperLink_]);
288 *latency += myRouter->myNodes_[myCoords[3] * numLinksperLink_]->latency();
290 if (hasLimiter_) { // limiter for sender
291 std::pair<surf::LinkImpl*, surf::LinkImpl*> info = privateLinks_.at(src->id() * linkCountPerNode_ + hasLoopback_);
292 route->link_list->push_back(info.first);
295 if (targetRouter != myRouter) {
297 // are we on a different group ?
298 if (targetRouter->group_ != currentRouter->group_) {
299 // go to the router of our group connected to this one.
300 if (currentRouter->blade_ != targetCoords[0]) {
301 // go to the nth router in our chassis
302 route->link_list->push_back(currentRouter->greenLinks_[targetCoords[0]]);
304 *latency += currentRouter->greenLinks_[targetCoords[0]]->latency();
305 currentRouter = routers_[myCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
306 myCoords[1] * numBladesPerChassis_ + targetCoords[0]];
309 if (currentRouter->chassis_ != 0) {
310 // go to the first chassis of our group
311 route->link_list->push_back(currentRouter->blackLinks_[0]);
313 *latency += currentRouter->blackLinks_[0]->latency();
314 currentRouter = routers_[myCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) + targetCoords[0]];
317 // go to destination group - the only optical hop
318 route->link_list->push_back(currentRouter->blueLinks_[0]);
320 *latency += currentRouter->blueLinks_[0]->latency();
321 currentRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) + myCoords[0]];
324 // same group, but same blade ?
325 if (targetRouter->blade_ != currentRouter->blade_) {
326 route->link_list->push_back(currentRouter->greenLinks_[targetCoords[2]]);
328 *latency += currentRouter->greenLinks_[targetCoords[2]]->latency();
329 currentRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) + targetCoords[2]];
332 // same blade, but same chassis ?
333 if (targetRouter->chassis_ != currentRouter->chassis_) {
334 route->link_list->push_back(currentRouter->blackLinks_[targetCoords[1]]);
336 *latency += currentRouter->blackLinks_[targetCoords[1]]->latency();
337 currentRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
338 targetCoords[1] * numBladesPerChassis_ + targetCoords[2]];
342 if (hasLimiter_) { // limiter for receiver
343 std::pair<surf::LinkImpl*, surf::LinkImpl*> info = privateLinks_.at(dst->id() * linkCountPerNode_ + hasLoopback_);
344 route->link_list->push_back(info.first);
347 // router->node local link
348 route->link_list->push_back(targetRouter->myNodes_[targetCoords[3] * numLinksperLink_ + numLinksperLink_ - 1]);
350 *latency += targetRouter->myNodes_[targetCoords[3] * numLinksperLink_ + numLinksperLink_ - 1]->latency();