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()
138 delete[] greenLinks_;
139 delete[] blackLinks_;
143 void DragonflyZone::generateRouters()
145 this->routers_ = new DragonflyRouter*[this->numGroups_ * this->numChassisPerGroup_ * this->numBladesPerChassis_];
147 for (unsigned int i = 0; i < this->numGroups_; i++) {
148 for (unsigned int j = 0; j < this->numChassisPerGroup_; j++) {
149 for (unsigned int k = 0; k < this->numBladesPerChassis_; k++) {
150 DragonflyRouter* router = new DragonflyRouter(i, j, k);
151 this->routers_[i * this->numChassisPerGroup_ * this->numBladesPerChassis_ + j * this->numBladesPerChassis_ +
158 void DragonflyZone::createLink(const std::string& id, int numlinks, surf::LinkImpl** linkup, surf::LinkImpl** linkdown)
162 LinkCreationArgs linkTemplate;
163 linkTemplate.bandwidth = this->cluster_->bw * numlinks;
164 linkTemplate.latency = this->cluster_->lat;
165 linkTemplate.policy = this->cluster_->sharing_policy; // sthg to do with that ?
166 linkTemplate.id = id;
167 sg_platf_new_link(&linkTemplate);
168 XBT_DEBUG("Generating link %s", id.c_str());
169 surf::LinkImpl* link;
171 if (this->cluster_->sharing_policy == SURF_LINK_FULLDUPLEX) {
172 tmpID = linkTemplate.id + "_UP";
173 link = surf::LinkImpl::byName(tmpID);
174 *linkup = link; // check link?
175 tmpID = linkTemplate.id + "_DOWN";
176 link = surf::LinkImpl::byName(tmpID);
177 *linkdown = link; // check link ?
179 link = surf::LinkImpl::byName(linkTemplate.id);
185 void DragonflyZone::generateLinks()
187 static int uniqueId = 0;
188 surf::LinkImpl* linkup;
189 surf::LinkImpl* linkdown;
191 unsigned int numRouters = this->numGroups_ * this->numChassisPerGroup_ * this->numBladesPerChassis_;
193 if (this->cluster_->sharing_policy == SURF_LINK_FULLDUPLEX)
194 numLinksperLink_ = 2;
196 // Links from routers to their local nodes.
197 for (unsigned int i = 0; i < numRouters; i++) {
198 // allocate structures
199 this->routers_[i]->myNodes_ = new surf::LinkImpl*[numLinksperLink_ * this->numNodesPerBlade_];
200 this->routers_[i]->greenLinks_ = new surf::LinkImpl*[this->numBladesPerChassis_];
201 this->routers_[i]->blackLinks_ = new surf::LinkImpl*[this->numChassisPerGroup_];
203 for (unsigned int j = 0; j < numLinksperLink_ * this->numNodesPerBlade_; j += numLinksperLink_) {
204 std::string id = "local_link_from_router_"+ std::to_string(i) + "_to_node_" +
205 std::to_string(j / numLinksperLink_) + "_" + std::to_string(uniqueId);
206 this->createLink(id, 1, &linkup, &linkdown);
208 if (this->cluster_->sharing_policy == SURF_LINK_FULLDUPLEX) {
209 this->routers_[i]->myNodes_[j] = linkup;
210 this->routers_[i]->myNodes_[j + 1] = linkdown;
212 this->routers_[i]->myNodes_[j] = linkup;
218 // Green links from routers to same chassis routers - alltoall
219 for (unsigned int i = 0; i < this->numGroups_ * this->numChassisPerGroup_; i++) {
220 for (unsigned int j = 0; j < this->numBladesPerChassis_; j++) {
221 for (unsigned int k = j + 1; k < this->numBladesPerChassis_; k++) {
222 std::string id = "green_link_in_chassis_" + std::to_string(i % numChassisPerGroup_) +"_between_routers_" +
223 std::to_string(j) + "_and_" + std::to_string(k) + "_" + std::to_string(uniqueId);
224 this->createLink(id, this->numLinksGreen_, &linkup, &linkdown);
226 this->routers_[i * numBladesPerChassis_ + j]->greenLinks_[k] = linkup;
227 this->routers_[i * numBladesPerChassis_ + k]->greenLinks_[j] = linkdown;
233 // Black links from routers to same group routers - alltoall
234 for (unsigned int i = 0; i < this->numGroups_; i++) {
235 for (unsigned int j = 0; j < this->numChassisPerGroup_; j++) {
236 for (unsigned int k = j + 1; k < this->numChassisPerGroup_; k++) {
237 for (unsigned int l = 0; l < this->numBladesPerChassis_; l++) {
238 std::string id = "black_link_in_group_" + std::to_string(i) + "_between_chassis_" + std::to_string(j) +
239 "_and_" + std::to_string(k) +"_blade_" + std::to_string(l) + "_" + std::to_string(uniqueId);
240 this->createLink(id, this->numLinksBlack_, &linkup, &linkdown);
242 this->routers_[i * numBladesPerChassis_ * numChassisPerGroup_ + j * numBladesPerChassis_ + l]
243 ->blackLinks_[k] = linkup;
244 this->routers_[i * numBladesPerChassis_ * numChassisPerGroup_ + k * numBladesPerChassis_ + l]
245 ->blackLinks_[j] = linkdown;
252 // Blue links between groups - Not all routers involved, only one per group is linked to others. Let's say router n of
253 // each group is linked to group n.
254 // FIXME: in reality blue links may be attached to several different routers
255 for (unsigned int i = 0; i < this->numGroups_; i++) {
256 for (unsigned int j = i + 1; j < this->numGroups_; j++) {
257 unsigned int routernumi = i * numBladesPerChassis_ * numChassisPerGroup_ + j;
258 unsigned int routernumj = j * numBladesPerChassis_ * numChassisPerGroup_ + i;
259 this->routers_[routernumi]->blueLinks_ = new surf::LinkImpl*;
260 this->routers_[routernumj]->blueLinks_ = new surf::LinkImpl*;
261 std::string id = "blue_link_between_group_"+ std::to_string(i) +"_and_" + std::to_string(j) +"_routers_" +
262 std::to_string(routernumi) + "_and_" + std::to_string(routernumj) + "_" + std::to_string(uniqueId);
263 this->createLink(id, this->numLinksBlue_, &linkup, &linkdown);
265 this->routers_[routernumi]->blueLinks_[0] = linkup;
266 this->routers_[routernumj]->blueLinks_[0] = linkdown;
272 void DragonflyZone::getLocalRoute(NetPoint* src, NetPoint* dst, sg_platf_route_cbarg_t route, double* latency)
274 // Minimal routing version.
275 // TODO : non-minimal random one, and adaptive ?
277 if (dst->isRouter() || src->isRouter())
280 XBT_VERB("dragonfly getLocalRoute from '%s'[%u] to '%s'[%u]", src->getCname(), src->id(), dst->getCname(), dst->id());
282 if ((src->id() == dst->id()) && hasLoopback_) {
283 std::pair<surf::LinkImpl*, surf::LinkImpl*> info = privateLinks_.at(nodePosition(src->id()));
285 route->link_list->push_back(info.first);
287 *latency += info.first->latency();
291 unsigned int myCoords[4];
292 rankId_to_coords(src->id(), &myCoords);
293 unsigned int targetCoords[4];
294 rankId_to_coords(dst->id(), &targetCoords);
295 XBT_DEBUG("src : %u group, %u chassis, %u blade, %u node", myCoords[0], myCoords[1], myCoords[2], myCoords[3]);
296 XBT_DEBUG("dst : %u group, %u chassis, %u blade, %u node", targetCoords[0], targetCoords[1], targetCoords[2],
299 DragonflyRouter* myRouter = routers_[myCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
300 myCoords[1] * numBladesPerChassis_ + myCoords[2]];
301 DragonflyRouter* targetRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
302 targetCoords[1] * numBladesPerChassis_ + targetCoords[2]];
303 DragonflyRouter* currentRouter = myRouter;
305 // node->router local link
306 route->link_list->push_back(myRouter->myNodes_[myCoords[3] * numLinksperLink_]);
308 *latency += myRouter->myNodes_[myCoords[3] * numLinksperLink_]->latency();
310 if (hasLimiter_) { // limiter for sender
311 std::pair<surf::LinkImpl*, surf::LinkImpl*> info = privateLinks_.at(nodePositionWithLoopback(src->id()));
312 route->link_list->push_back(info.first);
315 if (targetRouter != myRouter) {
317 // are we on a different group ?
318 if (targetRouter->group_ != currentRouter->group_) {
319 // go to the router of our group connected to this one.
320 if (currentRouter->blade_ != targetCoords[0]) {
321 // go to the nth router in our chassis
322 route->link_list->push_back(currentRouter->greenLinks_[targetCoords[0]]);
324 *latency += currentRouter->greenLinks_[targetCoords[0]]->latency();
325 currentRouter = routers_[myCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
326 myCoords[1] * numBladesPerChassis_ + targetCoords[0]];
329 if (currentRouter->chassis_ != 0) {
330 // go to the first chassis of our group
331 route->link_list->push_back(currentRouter->blackLinks_[0]);
333 *latency += currentRouter->blackLinks_[0]->latency();
334 currentRouter = routers_[myCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) + targetCoords[0]];
337 // go to destination group - the only optical hop
338 route->link_list->push_back(currentRouter->blueLinks_[0]);
340 *latency += currentRouter->blueLinks_[0]->latency();
341 currentRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) + myCoords[0]];
344 // same group, but same blade ?
345 if (targetRouter->blade_ != currentRouter->blade_) {
346 route->link_list->push_back(currentRouter->greenLinks_[targetCoords[2]]);
348 *latency += currentRouter->greenLinks_[targetCoords[2]]->latency();
349 currentRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) + targetCoords[2]];
352 // same blade, but same chassis ?
353 if (targetRouter->chassis_ != currentRouter->chassis_) {
354 route->link_list->push_back(currentRouter->blackLinks_[targetCoords[1]]);
356 *latency += currentRouter->blackLinks_[targetCoords[1]]->latency();
360 if (hasLimiter_) { // limiter for receiver
361 std::pair<surf::LinkImpl*, surf::LinkImpl*> info = privateLinks_.at(nodePositionWithLoopback(dst->id()));
362 route->link_list->push_back(info.first);
365 // router->node local link
366 route->link_list->push_back(targetRouter->myNodes_[targetCoords[3] * numLinksperLink_ + numLinksperLink_ - 1]);
368 *latency += targetRouter->myNodes_[targetCoords[3] * numLinksperLink_ + numLinksperLink_ - 1]->latency();