1 /* Copyright (c) 2014-2017. 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, std::string 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(ClusterCreationArgs* 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 if (parameters.size() != 4 || parameters.empty()) {
52 "Dragonfly are defined by the number of groups, chassis per groups, blades per chassis, nodes per blade");
55 // Blue network : number of groups, number of links between each group
56 boost::split(tmp, parameters[0], boost::is_any_of(","));
57 if (tmp.size() != 2) {
58 surf_parse_error("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
62 this->numGroups_ = std::stoi(tmp[0]);
63 } catch (std::invalid_argument& ia) {
64 throw std::invalid_argument(std::string("Invalid number of groups:") + tmp[0]);
68 this->numLinksBlue_ = std::stoi(tmp[1]);
69 } catch (std::invalid_argument& ia) {
70 throw std::invalid_argument(std::string("Invalid number of links for the blue level:") + tmp[1]);
72 // Black network : number of chassis/group, number of links between each router on the black network
73 boost::split(tmp, parameters[1], boost::is_any_of(","));
74 if (tmp.size() != 2) {
75 surf_parse_error("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
79 this->numChassisPerGroup_ = std::stoi(tmp[0]);
80 } catch (std::invalid_argument& ia) {
81 throw std::invalid_argument(std::string("Invalid number of groups:") + tmp[0]);
85 this->numLinksBlack_ = std::stoi(tmp[1]);
86 } catch (std::invalid_argument& ia) {
87 throw std::invalid_argument(std::string("Invalid number of links for the black level:") + tmp[1]);
90 // Green network : number of blades/chassis, number of links between each router on the green network
91 boost::split(tmp, parameters[2], boost::is_any_of(","));
92 if (tmp.size() != 2) {
93 surf_parse_error("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
97 this->numBladesPerChassis_ = std::stoi(tmp[0]);
98 } catch (std::invalid_argument& ia) {
99 throw std::invalid_argument(std::string("Invalid number of groups:") + tmp[0]);
103 this->numLinksGreen_ = std::stoi(tmp[1]);
104 } catch (std::invalid_argument& ia) {
105 throw std::invalid_argument(std::string("Invalid number of links for the green level:") + tmp[1]);
108 // The last part of topo_parameters should be the number of nodes per blade
110 this->numNodesPerBlade_ = std::stoi(parameters[3]);
111 } catch (std::invalid_argument& ia) {
112 throw std::invalid_argument(std::string("Last parameter is not the amount of nodes per blade:") + parameters[3]);
115 this->cluster_ = cluster;
119 * Generate the cluster once every node is created
121 void DragonflyZone::seal()
123 if (this->numNodesPerBlade_ == 0) {
127 this->generateRouters();
128 this->generateLinks();
131 DragonflyRouter::DragonflyRouter(int group, int chassis, int blade) : group_(group), chassis_(chassis), blade_(blade)
135 DragonflyRouter::~DragonflyRouter()
137 if (this->myNodes_ != nullptr)
139 if (this->greenLinks_ != nullptr)
140 xbt_free(greenLinks_);
141 if (this->blackLinks_ != nullptr)
142 xbt_free(blackLinks_);
143 if (this->blueLinks_ != nullptr)
144 xbt_free(blueLinks_);
147 void DragonflyZone::generateRouters()
149 this->routers_ = static_cast<DragonflyRouter**>(xbt_malloc0(this->numGroups_ * this->numChassisPerGroup_ *
150 this->numBladesPerChassis_ * sizeof(DragonflyRouter*)));
152 for (unsigned int i = 0; i < this->numGroups_; i++) {
153 for (unsigned int j = 0; j < this->numChassisPerGroup_; j++) {
154 for (unsigned int k = 0; k < this->numBladesPerChassis_; k++) {
155 DragonflyRouter* router = new DragonflyRouter(i, j, k);
156 this->routers_[i * this->numChassisPerGroup_ * this->numBladesPerChassis_ + j * this->numBladesPerChassis_ +
163 void DragonflyZone::createLink(const std::string& id, int numlinks, surf::LinkImpl** linkup, surf::LinkImpl** linkdown)
167 LinkCreationArgs linkTemplate;
168 linkTemplate.bandwidth = this->cluster_->bw * numlinks;
169 linkTemplate.latency = this->cluster_->lat;
170 linkTemplate.policy = this->cluster_->sharing_policy; // sthg to do with that ?
171 linkTemplate.id = id;
172 sg_platf_new_link(&linkTemplate);
173 XBT_DEBUG("Generating link %s", id.c_str());
174 surf::LinkImpl* link;
176 if (this->cluster_->sharing_policy == SURF_LINK_FULLDUPLEX) {
177 tmpID = linkTemplate.id + "_UP";
178 link = surf::LinkImpl::byName(tmpID);
179 *linkup = link; // check link?
180 tmpID = linkTemplate.id + "_DOWN";
181 link = surf::LinkImpl::byName(tmpID);
182 *linkdown = link; // check link ?
184 link = surf::LinkImpl::byName(linkTemplate.id);
190 void DragonflyZone::generateLinks()
192 static int uniqueId = 0;
193 surf::LinkImpl* linkup;
194 surf::LinkImpl* linkdown;
196 unsigned int numRouters = this->numGroups_ * this->numChassisPerGroup_ * this->numBladesPerChassis_;
198 if (this->cluster_->sharing_policy == SURF_LINK_FULLDUPLEX)
199 numLinksperLink_ = 2;
201 // Links from routers to their local nodes.
202 for (unsigned int i = 0; i < numRouters; i++) {
203 // allocate structures
204 this->routers_[i]->myNodes_ = static_cast<surf::LinkImpl**>(
205 xbt_malloc0(numLinksperLink_ * this->numNodesPerBlade_ * sizeof(surf::LinkImpl*)));
206 this->routers_[i]->greenLinks_ =
207 static_cast<surf::LinkImpl**>(xbt_malloc0(this->numBladesPerChassis_ * sizeof(surf::LinkImpl*)));
208 this->routers_[i]->blackLinks_ =
209 static_cast<surf::LinkImpl**>(xbt_malloc0(this->numChassisPerGroup_ * sizeof(surf::LinkImpl*)));
211 for (unsigned int j = 0; j < numLinksperLink_ * this->numNodesPerBlade_; j += numLinksperLink_) {
212 std::string id = "local_link_from_router_"+ std::to_string(i) + "_to_node_" +
213 std::to_string(j / numLinksperLink_) + "_" + std::to_string(uniqueId);
214 this->createLink(id, 1, &linkup, &linkdown);
216 if (this->cluster_->sharing_policy == SURF_LINK_FULLDUPLEX) {
217 this->routers_[i]->myNodes_[j] = linkup;
218 this->routers_[i]->myNodes_[j + 1] = linkdown;
220 this->routers_[i]->myNodes_[j] = linkup;
226 // Green links from routers to same chassis routers - alltoall
227 for (unsigned int i = 0; i < this->numGroups_ * this->numChassisPerGroup_; i++) {
228 for (unsigned int j = 0; j < this->numBladesPerChassis_; j++) {
229 for (unsigned int k = j + 1; k < this->numBladesPerChassis_; k++) {
230 std::string id = "green_link_in_chassis_" + std::to_string(i % numChassisPerGroup_) +"_between_routers_" +
231 std::to_string(j) + "_and_" + std::to_string(k) + "_" + std::to_string(uniqueId);
232 this->createLink(id, this->numLinksGreen_, &linkup, &linkdown);
234 this->routers_[i * numBladesPerChassis_ + j]->greenLinks_[k] = linkup;
235 this->routers_[i * numBladesPerChassis_ + k]->greenLinks_[j] = linkdown;
241 // Black links from routers to same group routers - alltoall
242 for (unsigned int i = 0; i < this->numGroups_; i++) {
243 for (unsigned int j = 0; j < this->numChassisPerGroup_; j++) {
244 for (unsigned int k = j + 1; k < this->numChassisPerGroup_; k++) {
245 for (unsigned int l = 0; l < this->numBladesPerChassis_; l++) {
246 std::string id = "black_link_in_group_" + std::to_string(i) + "_between_chassis_" + std::to_string(j) +
247 "_and_" + std::to_string(k) +"_blade_" + std::to_string(l) + "_" + std::to_string(uniqueId);
248 this->createLink(id, this->numLinksBlack_, &linkup, &linkdown);
250 this->routers_[i * numBladesPerChassis_ * numChassisPerGroup_ + j * numBladesPerChassis_ + l]
251 ->blackLinks_[k] = linkup;
252 this->routers_[i * numBladesPerChassis_ * numChassisPerGroup_ + k * numBladesPerChassis_ + l]
253 ->blackLinks_[j] = linkdown;
260 // Blue links between groups - Not all routers involved, only one per group is linked to others. Let's say router n of
261 // each group is linked to group n.
262 // FIXME: in reality blue links may be attached to several different routers
263 for (unsigned int i = 0; i < this->numGroups_; i++) {
264 for (unsigned int j = i + 1; j < this->numGroups_; j++) {
265 unsigned int routernumi = i * numBladesPerChassis_ * numChassisPerGroup_ + j;
266 unsigned int routernumj = j * numBladesPerChassis_ * numChassisPerGroup_ + i;
267 this->routers_[routernumi]->blueLinks_ = static_cast<surf::LinkImpl**>(xbt_malloc0(sizeof(surf::LinkImpl*)));
268 this->routers_[routernumj]->blueLinks_ = static_cast<surf::LinkImpl**>(xbt_malloc0(sizeof(surf::LinkImpl*)));
269 std::string id = "blue_link_between_group_"+ std::to_string(i) +"_and_" + std::to_string(j) +"_routers_" +
270 std::to_string(routernumi) + "_and_" + std::to_string(routernumj) + "_" + std::to_string(uniqueId);
271 this->createLink(id, this->numLinksBlue_, &linkup, &linkdown);
273 this->routers_[routernumi]->blueLinks_[0] = linkup;
274 this->routers_[routernumj]->blueLinks_[0] = linkdown;
280 void DragonflyZone::getLocalRoute(NetPoint* src, NetPoint* dst, sg_platf_route_cbarg_t route, double* latency)
282 // Minimal routing version.
283 // TODO : non-minimal random one, and adaptive ?
285 if (dst->isRouter() || src->isRouter())
288 XBT_VERB("dragonfly getLocalRout from '%s'[%u] to '%s'[%u]", src->name().c_str(), src->id(), dst->name().c_str(),
291 if ((src->id() == dst->id()) && hasLoopback_) {
292 std::pair<surf::LinkImpl*, surf::LinkImpl*> info = privateLinks_.at(nodePosition(src->id()));
294 route->link_list->push_back(info.first);
296 *latency += info.first->latency();
300 unsigned int myCoords[4];
301 rankId_to_coords(src->id(), &myCoords);
302 unsigned int targetCoords[4];
303 rankId_to_coords(dst->id(), &targetCoords);
304 XBT_DEBUG("src : %u group, %u chassis, %u blade, %u node", myCoords[0], myCoords[1], myCoords[2], myCoords[3]);
305 XBT_DEBUG("dst : %u group, %u chassis, %u blade, %u node", targetCoords[0], targetCoords[1], targetCoords[2],
308 DragonflyRouter* myRouter = routers_[myCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
309 myCoords[1] * numBladesPerChassis_ + myCoords[2]];
310 DragonflyRouter* targetRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
311 targetCoords[1] * numBladesPerChassis_ + targetCoords[2]];
312 DragonflyRouter* currentRouter = myRouter;
314 // node->router local link
315 route->link_list->push_back(myRouter->myNodes_[myCoords[3] * numLinksperLink_]);
317 *latency += myRouter->myNodes_[myCoords[3] * numLinksperLink_]->latency();
319 if (hasLimiter_) { // limiter for sender
320 std::pair<surf::LinkImpl*, surf::LinkImpl*> info = privateLinks_.at(nodePositionWithLoopback(src->id()));
321 route->link_list->push_back(info.first);
324 if (targetRouter != myRouter) {
326 // are we on a different group ?
327 if (targetRouter->group_ != currentRouter->group_) {
328 // go to the router of our group connected to this one.
329 if (currentRouter->blade_ != targetCoords[0]) {
330 // go to the nth router in our chassis
331 route->link_list->push_back(currentRouter->greenLinks_[targetCoords[0]]);
333 *latency += currentRouter->greenLinks_[targetCoords[0]]->latency();
334 currentRouter = routers_[myCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
335 myCoords[1] * numBladesPerChassis_ + targetCoords[0]];
338 if (currentRouter->chassis_ != 0) {
339 // go to the first chassis of our group
340 route->link_list->push_back(currentRouter->blackLinks_[0]);
342 *latency += currentRouter->blackLinks_[0]->latency();
343 currentRouter = routers_[myCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) + targetCoords[0]];
346 // go to destination group - the only optical hop
347 route->link_list->push_back(currentRouter->blueLinks_[0]);
349 *latency += currentRouter->blueLinks_[0]->latency();
350 currentRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) + myCoords[0]];
353 // same group, but same blade ?
354 if (targetRouter->blade_ != currentRouter->blade_) {
355 route->link_list->push_back(currentRouter->greenLinks_[targetCoords[2]]);
357 *latency += currentRouter->greenLinks_[targetCoords[2]]->latency();
358 currentRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) + targetCoords[2]];
361 // same blade, but same chassis ?
362 if (targetRouter->chassis_ != currentRouter->chassis_) {
363 route->link_list->push_back(currentRouter->blackLinks_[targetCoords[1]]);
365 *latency += currentRouter->blackLinks_[targetCoords[1]]->latency();
369 if (hasLimiter_) { // limiter for receiver
370 std::pair<surf::LinkImpl*, surf::LinkImpl*> info = privateLinks_.at(nodePositionWithLoopback(dst->id()));
371 route->link_list->push_back(info.first);
374 // router->node local link
375 route->link_list->push_back(targetRouter->myNodes_[targetCoords[3] * numLinksperLink_ + numLinksperLink_ - 1]);
377 *latency += targetRouter->myNodes_[targetCoords[3] * numLinksperLink_ + numLinksperLink_ - 1]->latency();