src/kernel/routing/DragonflyZone.cpp

   1 /* Copyright (c) 2014-2021. The SimGrid Team. All rights reserved.          */
   2
   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. */
   5
   6 #include "simgrid/kernel/routing/DragonflyZone.hpp"
   7 #include "simgrid/kernel/routing/NetPoint.hpp"
   8 #include "src/surf/network_interface.hpp"
   9
  10 #include <boost/algorithm/string/classification.hpp>
  11 #include <boost/algorithm/string/split.hpp>
  12 #include <numeric>
  13 #include <string>
  14
  15 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_route_cluster_dragonfly, surf_route_cluster, "Dragonfly Routing part of surf");
  16
  17 namespace simgrid {
  18 namespace kernel {
  19 namespace routing {
  20
  21 DragonflyZone::DragonflyZone(const std::string& name) : ClusterZone(name) {}
  22
  23 DragonflyZone::Coords DragonflyZone::rankId_to_coords(int rankId) const
  24 {
  25   // coords : group, chassis, blade, node
  26   Coords coords;
  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_;
  33   return coords;
  34 }
  35
  36 void DragonflyZone::rankId_to_coords(int rankId, unsigned int coords[4]) const // XBT_ATTRIB_DEPRECATED_v330
  37 {
  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;
  43 }
  44
  45 void DragonflyZone::set_link_characteristics(double bw, double lat, s4u::Link::SharingPolicy sharing_policy)
  46 {
  47   ClusterZone::set_link_characteristics(bw, lat, sharing_policy);
  48   if (sharing_policy == s4u::Link::SharingPolicy::SPLITDUPLEX)
  49     num_links_per_link_ = 2;
  50 }
  51
  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,
  54                                  unsigned int nodes)
  55 {
  56   num_groups_     = n_groups;
  57   num_links_blue_ = groups_links;
  58
  59   num_chassis_per_group_ = n_chassis;
  60   num_links_black_       = chassis_links;
  61
  62   num_blades_per_chassis_ = n_routers;
  63   num_links_green_        = routers_links;
  64
  65   num_nodes_per_blade_ = nodes;
  66 }
  67
  68 s4u::DragonflyParams DragonflyZone::parse_topo_parameters(const std::string& topo_parameters)
  69 {
  70   std::vector<std::string> parameters;
  71   std::vector<std::string> tmp;
  72   boost::split(parameters, topo_parameters, boost::is_any_of(";"));
  73
  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");
  76
  77   // Blue network : number of groups, number of links between each group
  78   boost::split(tmp, parameters[0], boost::is_any_of(","));
  79   if (tmp.size() != 2)
  80     xbt_die("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
  81
  82   unsigned int n_groups;
  83   try {
  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]);
  87   }
  88
  89   unsigned int n_blue;
  90   try {
  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]);
  94   }
  95
  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(","));
  98   if (tmp.size() != 2)
  99     xbt_die("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
 100
 101   unsigned int n_chassis;
 102   try {
 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]);
 106   }
 107
 108   unsigned int n_black;
 109   try {
 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]);
 113   }
 114
 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(","));
 117   if (tmp.size() != 2)
 118     xbt_die("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
 119
 120   unsigned int n_routers;
 121   try {
 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]);
 125   }
 126
 127   unsigned int n_green;
 128   try {
 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]);
 132   }
 133
 134   // The last part of topo_parameters should be the number of nodes per blade
 135   unsigned int n_nodes;
 136   try {
 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]);
 140   }
 141   return s4u::DragonflyParams({n_groups, n_blue}, {n_chassis, n_black}, {n_routers, n_green}, n_nodes);
 142 }
 143
 144 /* Generate the cluster once every node is created */
 145 void DragonflyZone::build_upper_levels(const s4u::ClusterCallbacks& set_callbacks)
 146 {
 147   generate_routers(set_callbacks);
 148   generate_links();
 149 }
 150
 151 void DragonflyZone::generate_routers(const s4u::ClusterCallbacks& set_callbacks)
 152 {
 153   int id = 0;
 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) {
 159       const auto* s4u_link =
 160           set_callbacks.limiter(get_iface(), {i, j, k, std::numeric_limits<unsigned int>::max()}, --id);
 161       if (s4u_link) {
 162         limiter = s4u_link->get_impl();
 163       }
 164     }
 165     return limiter;
 166   };
 167
 168   routers_.reserve(num_groups_ * num_chassis_per_group_ * num_blades_per_chassis_);
 169   for (unsigned int i = 0; i < num_groups_; i++) {
 170     for (unsigned int j = 0; j < num_chassis_per_group_; j++) {
 171       for (unsigned int k = 0; k < num_blades_per_chassis_; k++) {
 172         routers_.emplace_back(i, j, k, get_limiter(i, j, k));
 173       }
 174     }
 175   }
 176 }
 177
 178 void DragonflyZone::generate_link(const std::string& id, int numlinks, resource::LinkImpl** linkup,
 179                                   resource::LinkImpl** linkdown)
 180 {
 181   XBT_DEBUG("Generating link %s", id.c_str());
 182   *linkup   = nullptr;
 183   *linkdown = nullptr;
 184   if (get_link_sharing_policy() == s4u::Link::SharingPolicy::SPLITDUPLEX) {
 185     *linkup = create_link(id + "_UP", std::vector<double>{get_link_bandwidth() * numlinks})
 186                   ->set_latency(get_link_latency())
 187                   ->seal()
 188                   ->get_impl();
 189     *linkdown = create_link(id + "_DOWN", std::vector<double>{get_link_bandwidth() * numlinks})
 190                     ->set_latency(get_link_latency())
 191                     ->seal()
 192                     ->get_impl();
 193   } else {
 194     *linkup = create_link(id, std::vector<double>{get_link_bandwidth() * numlinks})
 195                   ->set_latency(get_link_latency())
 196                   ->seal()
 197                   ->get_impl();
 198     *linkdown = *linkup;
 199   }
 200 }
 201
 202 void DragonflyZone::generate_links()
 203 {
 204   static int uniqueId = 0;
 205   resource::LinkImpl* linkup;
 206   resource::LinkImpl* linkdown;
 207
 208   unsigned int numRouters = num_groups_ * num_chassis_per_group_ * num_blades_per_chassis_;
 209
 210   // Links from routers to their local nodes.
 211   for (unsigned int i = 0; i < numRouters; i++) {
 212     // allocate structures
 213     routers_[i].my_nodes_.resize(num_links_per_link_ * num_nodes_per_blade_);
 214     routers_[i].green_links_.resize(num_blades_per_chassis_);
 215     routers_[i].black_links_.resize(num_chassis_per_group_);
 216
 217     for (unsigned int j = 0; j < num_links_per_link_ * num_nodes_per_blade_; j += num_links_per_link_) {
 218       std::string id = "local_link_from_router_" + std::to_string(i) + "_to_node_" +
 219                        std::to_string(j / num_links_per_link_) + "_" + std::to_string(uniqueId);
 220       generate_link(id, 1, &linkup, &linkdown);
 221
 222       routers_[i].my_nodes_[j] = linkup;
 223       if (get_link_sharing_policy() == s4u::Link::SharingPolicy::SPLITDUPLEX)
 224         routers_[i].my_nodes_[j + 1] = linkdown;
 225
 226       uniqueId++;
 227     }
 228   }
 229
 230   // Green links from routers to same chassis routers - alltoall
 231   for (unsigned int i = 0; i < num_groups_ * num_chassis_per_group_; i++) {
 232     for (unsigned int j = 0; j < num_blades_per_chassis_; j++) {
 233       for (unsigned int k = j + 1; k < num_blades_per_chassis_; k++) {
 234         std::string id = "green_link_in_chassis_" + std::to_string(i % num_chassis_per_group_) + "_between_routers_" +
 235                          std::to_string(j) + "_and_" + std::to_string(k) + "_" + std::to_string(uniqueId);
 236         generate_link(id, num_links_green_, &linkup, &linkdown);
 237
 238         routers_[i * num_blades_per_chassis_ + j].green_links_[k] = linkup;
 239         routers_[i * num_blades_per_chassis_ + k].green_links_[j] = linkdown;
 240         uniqueId++;
 241       }
 242     }
 243   }
 244
 245   // Black links from routers to same group routers - alltoall
 246   for (unsigned int i = 0; i < num_groups_; i++) {
 247     for (unsigned int j = 0; j < num_chassis_per_group_; j++) {
 248       for (unsigned int k = j + 1; k < num_chassis_per_group_; k++) {
 249         for (unsigned int l = 0; l < num_blades_per_chassis_; l++) {
 250           std::string id = "black_link_in_group_" + std::to_string(i) + "_between_chassis_" + std::to_string(j) +
 251                            "_and_" + std::to_string(k) + "_blade_" + std::to_string(l) + "_" + std::to_string(uniqueId);
 252           generate_link(id, num_links_black_, &linkup, &linkdown);
 253
 254           routers_[i * num_blades_per_chassis_ * num_chassis_per_group_ + j * num_blades_per_chassis_ + l]
 255               .black_links_[k] = linkup;
 256           routers_[i * num_blades_per_chassis_ * num_chassis_per_group_ + k * num_blades_per_chassis_ + l]
 257               .black_links_[j] = linkdown;
 258           uniqueId++;
 259         }
 260       }
 261     }
 262   }
 263
 264   // Blue links between groups - Not all routers involved, only one per group is linked to others. Let's say router n of
 265   // each group is linked to group n.
 266   // FIXME: in reality blue links may be attached to several different routers
 267   for (unsigned int i = 0; i < num_groups_; i++) {
 268     for (unsigned int j = i + 1; j < num_groups_; j++) {
 269       unsigned int routernumi = i * num_blades_per_chassis_ * num_chassis_per_group_ + j;
 270       unsigned int routernumj = j * num_blades_per_chassis_ * num_chassis_per_group_ + i;
 271       std::string id = "blue_link_between_group_" + std::to_string(i) + "_and_" + std::to_string(j) + "_routers_" +
 272                        std::to_string(routernumi) + "_and_" + std::to_string(routernumj) + "_" +
 273                        std::to_string(uniqueId);
 274       generate_link(id, num_links_blue_, &linkup, &linkdown);
 275
 276       routers_[routernumi].blue_link_ = linkup;
 277       routers_[routernumj].blue_link_ = linkdown;
 278       uniqueId++;
 279     }
 280   }
 281 }
 282
 283 void DragonflyZone::get_local_route(NetPoint* src, NetPoint* dst, Route* route, double* latency)
 284 {
 285   // Minimal routing version.
 286   // TODO : non-minimal random one, and adaptive ?
 287
 288   if (dst->is_router() || src->is_router())
 289     return;
 290
 291   XBT_VERB("dragonfly getLocalRoute from '%s'[%u] to '%s'[%u]", src->get_cname(), src->id(), dst->get_cname(),
 292            dst->id());
 293
 294   if ((src->id() == dst->id()) && has_loopback()) {
 295     resource::LinkImpl* uplink = get_uplink_from(node_pos(src->id()));
 296
 297     route->link_list_.push_back(uplink);
 298     if (latency)
 299       *latency += uplink->get_latency();
 300     return;
 301   }
 302
 303   const auto myCoords     = rankId_to_coords(src->id());
 304   const auto targetCoords = rankId_to_coords(dst->id());
 305   XBT_DEBUG("src : %u group, %u chassis, %u blade, %u node", myCoords.group, myCoords.chassis, myCoords.blade,
 306             myCoords.node);
 307   XBT_DEBUG("dst : %u group, %u chassis, %u blade, %u node", targetCoords.group, targetCoords.chassis,
 308             targetCoords.blade, targetCoords.node);
 309
 310   DragonflyRouter* myRouter      = &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
 311                                         myCoords.chassis * num_blades_per_chassis_ + myCoords.blade];
 312   DragonflyRouter* targetRouter  = &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
 313                                             targetCoords.chassis * num_blades_per_chassis_ + targetCoords.blade];
 314   DragonflyRouter* currentRouter = myRouter;
 315
 316   if (has_limiter()) { // limiter for sender
 317     route->link_list_.push_back(get_uplink_from(node_pos_with_loopback(src->id())));
 318   }
 319
 320   // node->router local link
 321   route->link_list_.push_back(myRouter->my_nodes_[myCoords.node * num_links_per_link_]);
 322   if (latency)
 323     *latency += myRouter->my_nodes_[myCoords.node * num_links_per_link_]->get_latency();
 324
 325   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.group) {
 330         if (currentRouter->limiter_)
 331           route->link_list_.push_back(currentRouter->limiter_);
 332         // go to the nth router in our chassis
 333         route->link_list_.push_back(currentRouter->green_links_[targetCoords.group]);
 334         if (latency)
 335           *latency += currentRouter->green_links_[targetCoords.group]->get_latency();
 336         currentRouter = &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
 337                                   myCoords.chassis * num_blades_per_chassis_ + targetCoords.group];
 338       }
 339
 340       if (currentRouter->chassis_ != 0) {
 341         // go to the first chassis of our group
 342         if (currentRouter->limiter_)
 343           route->link_list_.push_back(currentRouter->limiter_);
 344         route->link_list_.push_back(currentRouter->black_links_[0]);
 345         if (latency)
 346           *latency += currentRouter->black_links_[0]->get_latency();
 347         currentRouter =
 348             &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + targetCoords.group];
 349       }
 350
 351       // go to destination group - the only optical hop
 352       route->link_list_.push_back(currentRouter->blue_link_);
 353       if (currentRouter->limiter_)
 354         route->link_list_.push_back(currentRouter->limiter_);
 355       if (latency)
 356         *latency += currentRouter->blue_link_->get_latency();
 357       currentRouter =
 358           &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + myCoords.group];
 359     }
 360
 361     // same group, but same blade ?
 362     if (targetRouter->blade_ != currentRouter->blade_) {
 363       if (currentRouter->limiter_)
 364         route->link_list_.push_back(currentRouter->limiter_);
 365       route->link_list_.push_back(currentRouter->green_links_[targetCoords.blade]);
 366       if (latency)
 367         *latency += currentRouter->green_links_[targetCoords.blade]->get_latency();
 368       currentRouter =
 369           &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + targetCoords.blade];
 370     }
 371
 372     // same blade, but same chassis ?
 373     if (targetRouter->chassis_ != currentRouter->chassis_) {
 374       if (currentRouter->limiter_)
 375         route->link_list_.push_back(currentRouter->limiter_);
 376       route->link_list_.push_back(currentRouter->black_links_[targetCoords.chassis]);
 377       if (latency)
 378         *latency += currentRouter->black_links_[targetCoords.chassis]->get_latency();
 379     }
 380   }
 381
 382   // router->node local link
 383   if (targetRouter->limiter_)
 384     route->link_list_.push_back(targetRouter->limiter_);
 385   route->link_list_.push_back(
 386       targetRouter->my_nodes_[targetCoords.node * num_links_per_link_ + num_links_per_link_ - 1]);
 387
 388   if (latency)
 389     *latency +=
 390         targetRouter->my_nodes_[targetCoords.node * num_links_per_link_ + num_links_per_link_ - 1]->get_latency();
 391
 392   if (has_limiter()) { // limiter for receiver
 393     route->link_list_.push_back(get_downlink_to(node_pos_with_loopback(dst->id())));
 394   }
 395
 396   // set gateways (if any)
 397   route->gw_src_ = get_gateway(src->id());
 398   route->gw_dst_ = get_gateway(dst->id());
 399 }
 400 } // namespace routing
 401 } // namespace kernel
 402
 403 namespace s4u {
 404 DragonflyParams::DragonflyParams(const std::pair<unsigned int, unsigned int>& groups,
 405                                  const std::pair<unsigned int, unsigned int>& chassis,
 406                                  const std::pair<unsigned int, unsigned int>& routers, unsigned int nodes)
 407     : groups(groups), chassis(chassis), routers(routers), nodes(nodes)
 408 {
 409   if (groups.first == 0)
 410     throw std::invalid_argument("Dragonfly: Invalid number of groups, must be > 0");
 411   if (groups.second == 0)
 412     throw std::invalid_argument("Dragonfly: Invalid number of blue (groups) links, must be > 0");
 413   if (chassis.first == 0)
 414     throw std::invalid_argument("Dragonfly: Invalid number of chassis, must be > 0");
 415   if (chassis.second == 0)
 416     throw std::invalid_argument("Dragonfly: Invalid number of black (chassis) links, must be > 0");
 417   if (routers.first == 0)
 418     throw std::invalid_argument("Dragonfly: Invalid number of routers, must be > 0");
 419   if (routers.second == 0)
 420     throw std::invalid_argument("Dragonfly: Invalid number of green (routers) links, must be > 0");
 421   if (nodes == 0)
 422     throw std::invalid_argument("Dragonfly: Invalid number of nodes, must be > 0");
 423 }
 424
 425 NetZone* create_dragonfly_zone(const std::string& name, const NetZone* parent, const DragonflyParams& params,
 426                                const ClusterCallbacks& set_callbacks, double bandwidth, double latency,
 427                                Link::SharingPolicy sharing_policy)
 428 {
 429   /* initial checks */
 430   if (bandwidth <= 0)
 431     throw std::invalid_argument("DragonflyZone: incorrect bandwidth for internode communication, bw=" +
 432                                 std::to_string(bandwidth));
 433   if (latency < 0)
 434     throw std::invalid_argument("DragonflyZone: incorrect latency for internode communication, lat=" +
 435                                 std::to_string(latency));
 436
 437   /* creating zone */
 438   auto* zone = new kernel::routing::DragonflyZone(name);
 439   zone->set_topology(params.groups.first, params.groups.second, params.chassis.first, params.chassis.second,
 440                      params.routers.first, params.routers.second, params.nodes);
 441   if (parent)
 442     zone->set_parent(parent->get_impl());
 443   zone->set_link_characteristics(bandwidth, latency, sharing_policy);
 444
 445   /* populating it */
 446   std::vector<unsigned int> dimensions = {params.groups.first, params.chassis.first, params.routers.first,
 447                                           params.nodes};
 448   int tot_elements                     = std::accumulate(dimensions.begin(), dimensions.end(), 1, std::multiplies<>());
 449   for (int i = 0; i < tot_elements; i++) {
 450     kernel::routing::NetPoint* netpoint;
 451     Link* limiter;
 452     Link* loopback;
 453     zone->fill_leaf_from_cb(i, dimensions, set_callbacks, &netpoint, &loopback, &limiter);
 454   }
 455   zone->build_upper_levels(set_callbacks);
 456
 457   return zone->get_iface();
 458 }
 459 } // namespace s4u
 460
 461 } // namespace simgrid