1 /* Copyright (c) 2014-2021. 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 "simgrid/kernel/routing/DragonflyZone.hpp"
7 #include "simgrid/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>
15 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_route_cluster_dragonfly, surf_route_cluster, "Dragonfly Routing part of surf");
21 DragonflyZone::DragonflyZone(const std::string& name) : ClusterBase(name) {}
23 DragonflyZone::Coords DragonflyZone::rankId_to_coords(unsigned long rankId) const
25 // coords : group, chassis, blade, node
27 coords.group = rankId / (num_chassis_per_group_ * num_blades_per_chassis_ * num_nodes_per_blade_);
28 rankId = rankId % (num_chassis_per_group_ * num_blades_per_chassis_ * num_nodes_per_blade_);
29 coords.chassis = rankId / (num_blades_per_chassis_ * num_nodes_per_blade_);
30 rankId = rankId % (num_blades_per_chassis_ * num_nodes_per_blade_);
31 coords.blade = rankId / num_nodes_per_blade_;
32 coords.node = rankId % num_nodes_per_blade_;
36 void DragonflyZone::rankId_to_coords(int rankId, unsigned int coords[4]) const // XBT_ATTRIB_DEPRECATED_v330
38 const auto s_coords = rankId_to_coords(rankId);
39 coords[0] = s_coords.group;
40 coords[1] = s_coords.chassis;
41 coords[2] = s_coords.blade;
42 coords[3] = s_coords.node;
45 void DragonflyZone::set_link_characteristics(double bw, double lat, s4u::Link::SharingPolicy sharing_policy)
47 ClusterBase::set_link_characteristics(bw, lat, sharing_policy);
48 if (sharing_policy == s4u::Link::SharingPolicy::SPLITDUPLEX)
49 num_links_per_link_ = 2;
52 void DragonflyZone::set_topology(unsigned int n_groups, unsigned int groups_links, unsigned int n_chassis,
53 unsigned int chassis_links, unsigned int n_routers, unsigned int routers_links,
56 num_groups_ = n_groups;
57 num_links_blue_ = groups_links;
59 num_chassis_per_group_ = n_chassis;
60 num_links_black_ = chassis_links;
62 num_blades_per_chassis_ = n_routers;
63 num_links_green_ = routers_links;
65 num_nodes_per_blade_ = nodes;
68 s4u::DragonflyParams DragonflyZone::parse_topo_parameters(const std::string& topo_parameters)
70 std::vector<std::string> parameters;
71 std::vector<std::string> tmp;
72 boost::split(parameters, topo_parameters, boost::is_any_of(";"));
74 if (parameters.size() != 4)
75 xbt_die("Dragonfly are defined by the number of groups, chassis per groups, blades per chassis, nodes per blade");
77 // Blue network : number of groups, number of links between each group
78 boost::split(tmp, parameters[0], boost::is_any_of(","));
80 xbt_die("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
82 unsigned int n_groups;
84 n_groups = std::stoi(tmp[0]);
85 } catch (const std::invalid_argument&) {
86 throw std::invalid_argument(std::string("Invalid number of groups:") + tmp[0]);
91 n_blue = std::stoi(tmp[1]);
92 } catch (const std::invalid_argument&) {
93 throw std::invalid_argument(std::string("Invalid number of links for the blue level:") + tmp[1]);
96 // Black network : number of chassis/group, number of links between each router on the black network
97 boost::split(tmp, parameters[1], boost::is_any_of(","));
99 xbt_die("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
101 unsigned int n_chassis;
103 n_chassis = std::stoi(tmp[0]);
104 } catch (const std::invalid_argument&) {
105 throw std::invalid_argument(std::string("Invalid number of chassis:") + tmp[0]);
108 unsigned int n_black;
110 n_black = std::stoi(tmp[1]);
111 } catch (const std::invalid_argument&) {
112 throw std::invalid_argument(std::string("Invalid number of links for the black level:") + tmp[1]);
115 // Green network : number of blades/chassis, number of links between each router on the green network
116 boost::split(tmp, parameters[2], boost::is_any_of(","));
118 xbt_die("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
120 unsigned int n_routers;
122 n_routers = std::stoi(tmp[0]);
123 } catch (const std::invalid_argument&) {
124 throw std::invalid_argument(std::string("Invalid number of routers:") + tmp[0]);
127 unsigned int n_green;
129 n_green = std::stoi(tmp[1]);
130 } catch (const std::invalid_argument&) {
131 throw std::invalid_argument(std::string("Invalid number of links for the green level:") + tmp[1]);
134 // The last part of topo_parameters should be the number of nodes per blade
135 unsigned int n_nodes;
137 n_nodes = std::stoi(parameters[3]);
138 } catch (const std::invalid_argument&) {
139 throw std::invalid_argument(std::string("Last parameter is not the amount of nodes per blade:") + parameters[3]);
141 return s4u::DragonflyParams({n_groups, n_blue}, {n_chassis, n_black}, {n_routers, n_green}, n_nodes);
144 /* Generate the cluster once every node is created */
145 void DragonflyZone::build_upper_levels(const s4u::ClusterCallbacks& set_callbacks)
147 generate_routers(set_callbacks);
151 void DragonflyZone::generate_routers(const s4u::ClusterCallbacks& set_callbacks)
153 unsigned long id = 2 * num_groups_ * num_chassis_per_group_ * num_blades_per_chassis_ * num_nodes_per_blade_;
154 /* get limiter for this router */
155 auto get_limiter = [this, &id, &set_callbacks](unsigned int i, unsigned int j,
156 unsigned int k) -> resource::LinkImpl* {
157 kernel::resource::LinkImpl* limiter = nullptr;
158 if (set_callbacks.limiter) {
160 const auto* s4u_link =
161 set_callbacks.limiter(get_iface(), {i, j, k, std::numeric_limits<unsigned int>::max()}, id);
163 limiter = s4u_link->get_impl();
169 routers_.reserve(static_cast<size_t>(num_groups_) * num_chassis_per_group_ * num_blades_per_chassis_);
170 for (unsigned int i = 0; i < num_groups_; i++) {
171 for (unsigned int j = 0; j < num_chassis_per_group_; j++) {
172 for (unsigned int k = 0; k < num_blades_per_chassis_; k++) {
173 routers_.emplace_back(i, j, k, get_limiter(i, j, k));
179 void DragonflyZone::generate_link(const std::string& id, int numlinks, resource::LinkImpl** linkup,
180 resource::LinkImpl** linkdown)
182 XBT_DEBUG("Generating link %s", id.c_str());
185 if (get_link_sharing_policy() == s4u::Link::SharingPolicy::SPLITDUPLEX) {
187 create_link(id + "_UP", {get_link_bandwidth() * numlinks})->set_latency(get_link_latency())->seal()->get_impl();
188 *linkdown = create_link(id + "_DOWN", {get_link_bandwidth() * numlinks})
189 ->set_latency(get_link_latency())
193 *linkup = create_link(id, {get_link_bandwidth() * numlinks})->set_latency(get_link_latency())->seal()->get_impl();
198 void DragonflyZone::generate_links()
200 static int uniqueId = 0;
201 resource::LinkImpl* linkup;
202 resource::LinkImpl* linkdown;
204 unsigned int numRouters = num_groups_ * num_chassis_per_group_ * num_blades_per_chassis_;
206 // Links from routers to their local nodes.
207 for (unsigned int i = 0; i < numRouters; i++) {
208 // allocate structures
209 routers_[i].my_nodes_.resize(static_cast<size_t>(num_links_per_link_) * num_nodes_per_blade_);
210 routers_[i].green_links_.resize(num_blades_per_chassis_);
211 routers_[i].black_links_.resize(num_chassis_per_group_);
213 for (unsigned int j = 0; j < num_links_per_link_ * num_nodes_per_blade_; j += num_links_per_link_) {
214 std::string id = "local_link_from_router_" + std::to_string(i) + "_to_node_" +
215 std::to_string(j / num_links_per_link_) + "_" + std::to_string(uniqueId);
216 generate_link(id, 1, &linkup, &linkdown);
218 routers_[i].my_nodes_[j] = linkup;
219 if (get_link_sharing_policy() == s4u::Link::SharingPolicy::SPLITDUPLEX)
220 routers_[i].my_nodes_[j + 1] = linkdown;
226 // Green links from routers to same chassis routers - alltoall
227 for (unsigned int i = 0; i < num_groups_ * num_chassis_per_group_; i++) {
228 for (unsigned int j = 0; j < num_blades_per_chassis_; j++) {
229 for (unsigned int k = j + 1; k < num_blades_per_chassis_; k++) {
230 std::string id = "green_link_in_chassis_" + std::to_string(i % num_chassis_per_group_) + "_between_routers_" +
231 std::to_string(j) + "_and_" + std::to_string(k) + "_" + std::to_string(uniqueId);
232 generate_link(id, num_links_green_, &linkup, &linkdown);
234 routers_[i * num_blades_per_chassis_ + j].green_links_[k] = linkup;
235 routers_[i * num_blades_per_chassis_ + k].green_links_[j] = linkdown;
241 // Black links from routers to same group routers - alltoall
242 for (unsigned int i = 0; i < num_groups_; i++) {
243 for (unsigned int j = 0; j < num_chassis_per_group_; j++) {
244 for (unsigned int k = j + 1; k < num_chassis_per_group_; k++) {
245 for (unsigned int l = 0; l < num_blades_per_chassis_; 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 generate_link(id, num_links_black_, &linkup, &linkdown);
250 routers_[i * num_blades_per_chassis_ * num_chassis_per_group_ + j * num_blades_per_chassis_ + l]
251 .black_links_[k] = linkup;
252 routers_[i * num_blades_per_chassis_ * num_chassis_per_group_ + k * num_blades_per_chassis_ + l]
253 .black_links_[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 < num_groups_; i++) {
264 for (unsigned int j = i + 1; j < num_groups_; j++) {
265 unsigned int routernumi = i * num_blades_per_chassis_ * num_chassis_per_group_ + j;
266 unsigned int routernumj = j * num_blades_per_chassis_ * num_chassis_per_group_ + i;
267 std::string id = "blue_link_between_group_" + std::to_string(i) + "_and_" + std::to_string(j) + "_routers_" +
268 std::to_string(routernumi) + "_and_" + std::to_string(routernumj) + "_" +
269 std::to_string(uniqueId);
270 generate_link(id, num_links_blue_, &linkup, &linkdown);
272 routers_[routernumi].blue_link_ = linkup;
273 routers_[routernumj].blue_link_ = linkdown;
279 void DragonflyZone::get_local_route(const NetPoint* src, const NetPoint* dst, Route* route, double* latency)
281 // Minimal routing version.
282 // TODO : non-minimal random one, and adaptive ?
284 if (dst->is_router() || src->is_router())
287 XBT_VERB("dragonfly getLocalRoute from '%s'[%lu] to '%s'[%lu]", src->get_cname(), src->id(), dst->get_cname(),
290 if ((src->id() == dst->id()) && has_loopback()) {
291 resource::LinkImpl* uplink = get_uplink_from(node_pos(src->id()));
293 add_link_latency(route->link_list_, uplink, latency);
297 const auto myCoords = rankId_to_coords(src->id());
298 const auto targetCoords = rankId_to_coords(dst->id());
299 XBT_DEBUG("src : %lu group, %lu chassis, %lu blade, %lu node", myCoords.group, myCoords.chassis, myCoords.blade,
301 XBT_DEBUG("dst : %lu group, %lu chassis, %lu blade, %lu node", targetCoords.group, targetCoords.chassis,
302 targetCoords.blade, targetCoords.node);
304 DragonflyRouter* myRouter = &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
305 myCoords.chassis * num_blades_per_chassis_ + myCoords.blade];
306 DragonflyRouter* targetRouter = &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
307 targetCoords.chassis * num_blades_per_chassis_ + targetCoords.blade];
308 DragonflyRouter* currentRouter = myRouter;
310 if (has_limiter()) { // limiter for sender
311 route->link_list_.push_back(get_uplink_from(node_pos_with_loopback(src->id())));
314 // node->router local link
315 add_link_latency(route->link_list_, myRouter->my_nodes_[static_cast<size_t>(myCoords.node) * num_links_per_link_],
318 if (targetRouter != myRouter) {
319 // are we on a different group ?
320 if (targetRouter->group_ != currentRouter->group_) {
321 // go to the router of our group connected to this one.
322 if (currentRouter->blade_ != targetCoords.group) {
323 if (currentRouter->limiter_)
324 route->link_list_.push_back(currentRouter->limiter_);
325 // go to the nth router in our chassis
326 add_link_latency(route->link_list_, currentRouter->green_links_[targetCoords.group], latency);
327 currentRouter = &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
328 myCoords.chassis * num_blades_per_chassis_ + targetCoords.group];
331 if (currentRouter->chassis_ != 0) {
332 // go to the first chassis of our group
333 if (currentRouter->limiter_)
334 route->link_list_.push_back(currentRouter->limiter_);
335 add_link_latency(route->link_list_, currentRouter->black_links_[0], latency);
337 &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + targetCoords.group];
340 // go to destination group - the only optical hop
341 add_link_latency(route->link_list_, currentRouter->blue_link_, latency);
342 if (currentRouter->limiter_)
343 route->link_list_.push_back(currentRouter->limiter_);
345 &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + myCoords.group];
348 // same group, but same blade ?
349 if (targetRouter->blade_ != currentRouter->blade_) {
350 if (currentRouter->limiter_)
351 route->link_list_.push_back(currentRouter->limiter_);
352 add_link_latency(route->link_list_, currentRouter->green_links_[targetCoords.blade], latency);
354 &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + targetCoords.blade];
357 // same blade, but same chassis ?
358 if (targetRouter->chassis_ != currentRouter->chassis_) {
359 if (currentRouter->limiter_)
360 route->link_list_.push_back(currentRouter->limiter_);
361 add_link_latency(route->link_list_, currentRouter->black_links_[targetCoords.chassis], latency);
365 // router->node local link
366 if (targetRouter->limiter_)
367 route->link_list_.push_back(targetRouter->limiter_);
368 add_link_latency(route->link_list_,
369 targetRouter->my_nodes_[targetCoords.node * num_links_per_link_ + num_links_per_link_ - 1], latency);
371 if (has_limiter()) { // limiter for receiver
372 route->link_list_.push_back(get_downlink_to(node_pos_with_loopback(dst->id())));
375 // set gateways (if any)
376 route->gw_src_ = get_gateway(src->id());
377 route->gw_dst_ = get_gateway(dst->id());
379 } // namespace routing
380 } // namespace kernel
383 DragonflyParams::DragonflyParams(const std::pair<unsigned int, unsigned int>& groups,
384 const std::pair<unsigned int, unsigned int>& chassis,
385 const std::pair<unsigned int, unsigned int>& routers, unsigned int nodes)
386 : groups(groups), chassis(chassis), routers(routers), nodes(nodes)
388 if (groups.first == 0)
389 throw std::invalid_argument("Dragonfly: Invalid number of groups, must be > 0");
390 if (groups.second == 0)
391 throw std::invalid_argument("Dragonfly: Invalid number of blue (groups) links, must be > 0");
392 if (chassis.first == 0)
393 throw std::invalid_argument("Dragonfly: Invalid number of chassis, must be > 0");
394 if (chassis.second == 0)
395 throw std::invalid_argument("Dragonfly: Invalid number of black (chassis) links, must be > 0");
396 if (routers.first == 0)
397 throw std::invalid_argument("Dragonfly: Invalid number of routers, must be > 0");
398 if (routers.second == 0)
399 throw std::invalid_argument("Dragonfly: Invalid number of green (routers) links, must be > 0");
401 throw std::invalid_argument("Dragonfly: Invalid number of nodes, must be > 0");
404 NetZone* create_dragonfly_zone(const std::string& name, const NetZone* parent, const DragonflyParams& params,
405 const ClusterCallbacks& set_callbacks, double bandwidth, double latency,
406 Link::SharingPolicy sharing_policy)
410 throw std::invalid_argument("DragonflyZone: incorrect bandwidth for internode communication, bw=" +
411 std::to_string(bandwidth));
413 throw std::invalid_argument("DragonflyZone: incorrect latency for internode communication, lat=" +
414 std::to_string(latency));
417 auto* zone = new kernel::routing::DragonflyZone(name);
418 zone->set_topology(params.groups.first, params.groups.second, params.chassis.first, params.chassis.second,
419 params.routers.first, params.routers.second, params.nodes);
421 zone->set_parent(parent->get_impl());
422 zone->set_link_characteristics(bandwidth, latency, sharing_policy);
425 std::vector<unsigned long> dimensions = {params.groups.first, params.chassis.first, params.routers.first,
427 int tot_elements = std::accumulate(dimensions.begin(), dimensions.end(), 1, std::multiplies<>());
428 for (int i = 0; i < tot_elements; i++) {
429 kernel::routing::NetPoint* netpoint;
432 zone->fill_leaf_from_cb(i, dimensions, set_callbacks, &netpoint, &loopback, &limiter);
434 zone->build_upper_levels(set_callbacks);
435 return zone->get_iface();
439 } // namespace simgrid