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/NetCard.hpp"
8 #include "src/surf/network_interface.hpp"
10 #include <boost/algorithm/string/classification.hpp>
11 #include <boost/algorithm/string/split.hpp>
13 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_route_cluster_dragonfly, surf_route_cluster, "Dragonfly Routing part of surf");
19 DragonflyZone::DragonflyZone(NetZone* father, const char* name) : ClusterZone(father, name)
23 DragonflyZone::~DragonflyZone()
25 if (this->routers_ != nullptr) {
26 for (unsigned int i = 0; i < this->numGroups_ * this->numChassisPerGroup_ * this->numBladesPerChassis_; i++)
32 unsigned int* DragonflyZone::rankId_to_coords(int rankId)
34 // coords : group, chassis, blade, node
35 unsigned int* coords = (unsigned int*)malloc(4 * sizeof(unsigned int));
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_;
46 void DragonflyZone::parse_specific_arguments(sg_platf_cluster_cbarg_t cluster)
48 std::vector<std::string> parameters;
49 std::vector<std::string> tmp;
50 boost::split(parameters, cluster->topo_parameters, boost::is_any_of(";"));
52 // TODO : we have to check for zeros and negative numbers, or it might crash
53 if (parameters.size() != 4) {
55 "Dragonfly are defined by the number of groups, chassis per groups, blades per chassis, nodes per blade");
58 // Blue network : number of groups, number of links between each group
59 boost::split(tmp, parameters[0], boost::is_any_of(","));
60 if (tmp.size() != 2) {
61 surf_parse_error("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
64 this->numGroups_ = xbt_str_parse_int(tmp[0].c_str(), "Invalid number of groups: %s");
65 this->numLinksBlue_ = xbt_str_parse_int(tmp[1].c_str(), "Invalid number of links for the blue level: %s");
67 // Black network : number of chassis/group, number of links between each router on the black network
68 boost::split(tmp, parameters[1], boost::is_any_of(","));
69 if (tmp.size() != 2) {
70 surf_parse_error("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
73 this->numChassisPerGroup_ = xbt_str_parse_int(tmp[0].c_str(), "Invalid number of groups: %s");
74 this->numLinksBlack_ = xbt_str_parse_int(tmp[1].c_str(), "Invalid number of links for the black level: %s");
76 // Green network : number of blades/chassis, number of links between each router on the green network
77 boost::split(tmp, parameters[2], boost::is_any_of(","));
78 if (tmp.size() != 2) {
79 surf_parse_error("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
82 this->numBladesPerChassis_ = xbt_str_parse_int(tmp[0].c_str(), "Invalid number of groups: %s");
83 this->numLinksGreen_ = xbt_str_parse_int(tmp[1].c_str(), "Invalid number of links for the green level: %s");
85 // The last part of topo_parameters should be the number of nodes per blade
86 this->numNodesPerBlade_ =
87 xbt_str_parse_int(parameters[3].c_str(), "Last parameter is not the amount of nodes per blade: %s");
88 this->cluster_ = cluster;
92 * Generate the cluster once every node is created
94 void DragonflyZone::seal()
96 if (this->numNodesPerBlade_ == 0) {
100 this->generateRouters();
101 this->generateLinks();
104 DragonflyRouter::DragonflyRouter(int group, int chassis, int blade) : group_(group), chassis_(chassis), blade_(blade)
108 DragonflyRouter::~DragonflyRouter()
110 if (this->myNodes_ != nullptr)
112 if (this->greenLinks_ != nullptr)
113 xbt_free(greenLinks_);
114 if (this->blackLinks_ != nullptr)
115 xbt_free(blackLinks_);
116 if (this->blueLinks_ != nullptr)
117 xbt_free(blueLinks_);
120 void DragonflyZone::generateRouters()
122 this->routers_ = static_cast<DragonflyRouter**>(xbt_malloc0(this->numGroups_ * this->numChassisPerGroup_ *
123 this->numBladesPerChassis_ * sizeof(DragonflyRouter*)));
125 for (unsigned int i = 0; i < this->numGroups_; i++) {
126 for (unsigned int j = 0; j < this->numChassisPerGroup_; j++) {
127 for (unsigned int k = 0; k < this->numBladesPerChassis_; k++) {
128 DragonflyRouter* router = new DragonflyRouter(i, j, k);
129 this->routers_[i * this->numChassisPerGroup_ * this->numBladesPerChassis_ + j * this->numBladesPerChassis_ +
136 void DragonflyZone::createLink(char* id, int numlinks, Link** linkup, Link** linkdown)
140 s_sg_platf_link_cbarg_t linkTemplate;
141 memset(&linkTemplate, 0, sizeof(linkTemplate));
142 linkTemplate.bandwidth = this->cluster_->bw * numlinks;
143 linkTemplate.latency = this->cluster_->lat;
144 linkTemplate.policy = this->cluster_->sharing_policy; // sthg to do with that ?
145 linkTemplate.id = id;
146 sg_platf_new_link(&linkTemplate);
147 XBT_DEBUG("Generating link %s", id);
150 if (this->cluster_->sharing_policy == SURF_LINK_FULLDUPLEX) {
151 tmpID = std::string(linkTemplate.id) + "_UP";
152 link = Link::byName(tmpID.c_str());
153 *linkup = link; // check link?
154 tmpID = std::string(linkTemplate.id) + "_DOWN";
155 link = Link::byName(tmpID.c_str());
156 *linkdown = link; // check link ?
158 link = Link::byName(linkTemplate.id);
163 free((void*)linkTemplate.id);
166 void DragonflyZone::generateLinks()
169 static int uniqueId = 0;
174 unsigned int numRouters = this->numGroups_ * this->numChassisPerGroup_ * this->numBladesPerChassis_;
176 if (this->cluster_->sharing_policy == SURF_LINK_FULLDUPLEX)
177 numLinksperLink_ = 2;
179 // Links from routers to their local nodes.
180 for (unsigned int i = 0; i < numRouters; i++) {
181 // allocate structures
182 this->routers_[i]->myNodes_ =
183 static_cast<Link**>(xbt_malloc0(numLinksperLink_ * this->numNodesPerBlade_ * sizeof(Link*)));
184 this->routers_[i]->greenLinks_ = static_cast<Link**>(xbt_malloc0(this->numBladesPerChassis_ * sizeof(Link*)));
185 this->routers_[i]->blackLinks_ = static_cast<Link**>(xbt_malloc0(this->numChassisPerGroup_ * sizeof(Link*)));
187 for (unsigned int j = 0; j < numLinksperLink_ * this->numNodesPerBlade_; j += numLinksperLink_) {
188 id = bprintf("local_link_from_router_%d_to_node_%d_%d", i, j / numLinksperLink_, uniqueId);
189 this->createLink(id, 1, &linkup, &linkdown);
190 if (this->cluster_->sharing_policy == SURF_LINK_FULLDUPLEX) {
191 this->routers_[i]->myNodes_[j] = linkup;
192 this->routers_[i]->myNodes_[j + 1] = linkdown;
194 this->routers_[i]->myNodes_[j] = linkup;
200 // Green links from routers to same chassis routers - alltoall
201 for (unsigned int i = 0; i < this->numGroups_ * this->numChassisPerGroup_; i++) {
202 for (unsigned int j = 0; j < this->numBladesPerChassis_; j++) {
203 for (unsigned int k = j + 1; k < this->numBladesPerChassis_; k++) {
204 id = bprintf("green_link_in_chassis_%d_between_routers_%d_and_%d_%d", i % numChassisPerGroup_, j, k, uniqueId);
205 this->createLink(id, this->numLinksGreen_, &linkup, &linkdown);
206 this->routers_[i * numBladesPerChassis_ + j]->greenLinks_[k] = linkup;
207 this->routers_[i * numBladesPerChassis_ + k]->greenLinks_[j] = linkdown;
213 // Black links from routers to same group routers - alltoall
214 for (unsigned int i = 0; i < this->numGroups_; i++) {
215 for (unsigned int j = 0; j < this->numChassisPerGroup_; j++) {
216 for (unsigned int k = j + 1; k < this->numChassisPerGroup_; k++) {
217 for (unsigned int l = 0; l < this->numBladesPerChassis_; l++) {
218 id = bprintf("black_link_in_group_%d_between_chassis_%d_and_%d_blade_%d_%d", i, j, k, l, uniqueId);
219 this->createLink(id, this->numLinksBlack_, &linkup, &linkdown);
220 this->routers_[i * numBladesPerChassis_ * numChassisPerGroup_ + j * numBladesPerChassis_ + l]
221 ->blackLinks_[k] = linkup;
222 this->routers_[i * numBladesPerChassis_ * numChassisPerGroup_ + k * numBladesPerChassis_ + l]
223 ->blackLinks_[j] = linkdown;
230 // Blue links between groups - Not all routers involved, only one per group is linked to others. Let's say router n of
231 // each group is linked to group n.
232 // FIXME: in reality blue links may be attached to several different routers
233 for (unsigned int i = 0; i < this->numGroups_; i++) {
234 for (unsigned int j = i + 1; j < this->numGroups_; j++) {
235 unsigned int routernumi = i * numBladesPerChassis_ * numChassisPerGroup_ + j;
236 unsigned int routernumj = j * numBladesPerChassis_ * numChassisPerGroup_ + i;
237 this->routers_[routernumi]->blueLinks_ = static_cast<Link**>(xbt_malloc0(sizeof(Link*)));
238 this->routers_[routernumj]->blueLinks_ = static_cast<Link**>(xbt_malloc0(sizeof(Link*)));
239 id = bprintf("blue_link_between_group_%d_and_%d_routers_%d_and_%d_%d", i, j, routernumi, routernumj, uniqueId);
240 this->createLink(id, this->numLinksBlue_, &linkup, &linkdown);
241 this->routers_[routernumi]->blueLinks_[0] = linkup;
242 this->routers_[routernumj]->blueLinks_[0] = linkdown;
248 void DragonflyZone::getLocalRoute(NetCard* src, NetCard* dst, sg_platf_route_cbarg_t route, double* latency)
250 // Minimal routing version.
251 // TODO : non-minimal random one, and adaptive ?
253 if (dst->isRouter() || src->isRouter())
256 XBT_VERB("dragonfly getLocalRout from '%s'[%d] to '%s'[%d]", src->name().c_str(), src->id(), dst->name().c_str(),
259 if ((src->id() == dst->id()) && hasLoopback_) {
260 std::pair<Link*, Link*> info = privateLinks_.at(src->id() * linkCountPerNode_);
262 route->link_list->push_back(info.first);
264 *latency += info.first->latency();
268 unsigned int* myCoords = rankId_to_coords(src->id());
269 unsigned int* targetCoords = rankId_to_coords(dst->id());
270 XBT_DEBUG("src : %u group, %u chassis, %u blade, %u node", myCoords[0], myCoords[1], myCoords[2], myCoords[3]);
271 XBT_DEBUG("dst : %u group, %u chassis, %u blade, %u node", targetCoords[0], targetCoords[1], targetCoords[2],
274 DragonflyRouter* myRouter = routers_[myCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
275 myCoords[1] * numBladesPerChassis_ + myCoords[2]];
276 DragonflyRouter* targetRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
277 targetCoords[1] * numBladesPerChassis_ + targetCoords[2]];
278 DragonflyRouter* currentRouter = myRouter;
280 // node->router local link
281 route->link_list->push_back(myRouter->myNodes_[myCoords[3] * numLinksperLink_]);
283 *latency += myRouter->myNodes_[myCoords[3] * numLinksperLink_]->latency();
285 if (hasLimiter_) { // limiter for sender
286 std::pair<Link*, Link*> info = privateLinks_.at(src->id() * linkCountPerNode_ + hasLoopback_);
287 route->link_list->push_back(info.first);
290 if (targetRouter != myRouter) {
292 // are we on a different group ?
293 if (targetRouter->group_ != currentRouter->group_) {
294 // go to the router of our group connected to this one.
295 if (currentRouter->blade_ != targetCoords[0]) {
296 // go to the nth router in our chassis
297 route->link_list->push_back(currentRouter->greenLinks_[targetCoords[0]]);
299 *latency += currentRouter->greenLinks_[targetCoords[0]]->latency();
300 currentRouter = routers_[myCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
301 myCoords[1] * numBladesPerChassis_ + targetCoords[0]];
304 if (currentRouter->chassis_ != 0) {
305 // go to the first chassis of our group
306 route->link_list->push_back(currentRouter->blackLinks_[0]);
308 *latency += currentRouter->blackLinks_[0]->latency();
309 currentRouter = routers_[myCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) + targetCoords[0]];
312 // go to destination group - the only optical hop
313 route->link_list->push_back(currentRouter->blueLinks_[0]);
315 *latency += currentRouter->blueLinks_[0]->latency();
316 currentRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) + myCoords[0]];
319 // same group, but same blade ?
320 if (targetRouter->blade_ != currentRouter->blade_) {
321 route->link_list->push_back(currentRouter->greenLinks_[targetCoords[2]]);
323 *latency += currentRouter->greenLinks_[targetCoords[2]]->latency();
324 currentRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) + targetCoords[2]];
327 // same blade, but same chassis ?
328 if (targetRouter->chassis_ != currentRouter->chassis_) {
329 route->link_list->push_back(currentRouter->blackLinks_[targetCoords[1]]);
331 *latency += currentRouter->blackLinks_[targetCoords[1]]->latency();
332 currentRouter = routers_[targetCoords[0] * (numChassisPerGroup_ * numBladesPerChassis_) +
333 targetCoords[1] * numBladesPerChassis_ + targetCoords[2]];
337 if (hasLimiter_) { // limiter for receiver
338 std::pair<Link*, Link*> info = privateLinks_.at(dst->id() * linkCountPerNode_ + hasLoopback_);
339 route->link_list->push_back(info.first);
342 // router->node local link
343 route->link_list->push_back(targetRouter->myNodes_[targetCoords[3] * numLinksperLink_ + numLinksperLink_ - 1]);
345 *latency += targetRouter->myNodes_[targetCoords[3] * numLinksperLink_ + numLinksperLink_ - 1]->latency();
348 xbt_free(targetCoords);