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   routers_.reserve(num_groups_ * num_chassis_per_group_ * num_blades_per_chassis_);
 155   for (unsigned int i = 0; i < num_groups_; i++) {
 156     for (unsigned int j = 0; j < num_chassis_per_group_; j++) {
 157       for (unsigned int k = 0; k < num_blades_per_chassis_; k++) {
 158         resource::LinkImpl* limiter = nullptr;
 159         if (set_callbacks.limiter) {
 160           auto* s4u_link =
 161               set_callbacks.limiter(get_iface(), {i, j, k, std::numeric_limits<unsigned int>::max()}, --id);
 162           if (s4u_link)
 163             limiter = s4u_link->get_impl();
 164         }
 165         routers_.emplace_back(i, j, k, limiter);
 166       }
 167     }
 168   }
 169 }
 170
 171 void DragonflyZone::generate_link(const std::string& id, int numlinks, resource::LinkImpl** linkup,
 172                                   resource::LinkImpl** linkdown)
 173 {
 174   XBT_DEBUG("Generating link %s", id.c_str());
 175   *linkup   = nullptr;
 176   *linkdown = nullptr;
 177   if (get_link_sharing_policy() == s4u::Link::SharingPolicy::SPLITDUPLEX) {
 178     *linkup = create_link(id + "_UP", std::vector<double>{get_link_bandwidth() * numlinks})
 179                   ->set_latency(get_link_latency())
 180                   ->seal()
 181                   ->get_impl();
 182     *linkdown = create_link(id + "_DOWN", std::vector<double>{get_link_bandwidth() * numlinks})
 183                     ->set_latency(get_link_latency())
 184                     ->seal()
 185                     ->get_impl();
 186   } else {
 187     *linkup = create_link(id, std::vector<double>{get_link_bandwidth() * numlinks})
 188                   ->set_latency(get_link_latency())
 189                   ->seal()
 190                   ->get_impl();
 191     *linkdown = *linkup;
 192   }
 193 }
 194
 195 void DragonflyZone::generate_links()
 196 {
 197   static int uniqueId = 0;
 198   resource::LinkImpl* linkup;
 199   resource::LinkImpl* linkdown;
 200
 201   unsigned int numRouters = num_groups_ * num_chassis_per_group_ * num_blades_per_chassis_;
 202
 203   // Links from routers to their local nodes.
 204   for (unsigned int i = 0; i < numRouters; i++) {
 205     // allocate structures
 206     routers_[i].my_nodes_.resize(num_links_per_link_ * num_nodes_per_blade_);
 207     routers_[i].green_links_.resize(num_blades_per_chassis_);
 208     routers_[i].black_links_.resize(num_chassis_per_group_);
 209
 210     for (unsigned int j = 0; j < num_links_per_link_ * num_nodes_per_blade_; j += num_links_per_link_) {
 211       std::string id = "local_link_from_router_" + std::to_string(i) + "_to_node_" +
 212                        std::to_string(j / num_links_per_link_) + "_" + std::to_string(uniqueId);
 213       generate_link(id, 1, &linkup, &linkdown);
 214
 215       routers_[i].my_nodes_[j] = linkup;
 216       if (get_link_sharing_policy() == s4u::Link::SharingPolicy::SPLITDUPLEX)
 217         routers_[i].my_nodes_[j + 1] = linkdown;
 218
 219       uniqueId++;
 220     }
 221   }
 222
 223   // Green links from routers to same chassis routers - alltoall
 224   for (unsigned int i = 0; i < num_groups_ * num_chassis_per_group_; i++) {
 225     for (unsigned int j = 0; j < num_blades_per_chassis_; j++) {
 226       for (unsigned int k = j + 1; k < num_blades_per_chassis_; k++) {
 227         std::string id = "green_link_in_chassis_" + std::to_string(i % num_chassis_per_group_) + "_between_routers_" +
 228                          std::to_string(j) + "_and_" + std::to_string(k) + "_" + std::to_string(uniqueId);
 229         generate_link(id, num_links_green_, &linkup, &linkdown);
 230
 231         routers_[i * num_blades_per_chassis_ + j].green_links_[k] = linkup;
 232         routers_[i * num_blades_per_chassis_ + k].green_links_[j] = linkdown;
 233         uniqueId++;
 234       }
 235     }
 236   }
 237
 238   // Black links from routers to same group routers - alltoall
 239   for (unsigned int i = 0; i < num_groups_; i++) {
 240     for (unsigned int j = 0; j < num_chassis_per_group_; j++) {
 241       for (unsigned int k = j + 1; k < num_chassis_per_group_; k++) {
 242         for (unsigned int l = 0; l < num_blades_per_chassis_; l++) {
 243           std::string id = "black_link_in_group_" + std::to_string(i) + "_between_chassis_" + std::to_string(j) +
 244                            "_and_" + std::to_string(k) + "_blade_" + std::to_string(l) + "_" + std::to_string(uniqueId);
 245           generate_link(id, num_links_black_, &linkup, &linkdown);
 246
 247           routers_[i * num_blades_per_chassis_ * num_chassis_per_group_ + j * num_blades_per_chassis_ + l]
 248               .black_links_[k] = linkup;
 249           routers_[i * num_blades_per_chassis_ * num_chassis_per_group_ + k * num_blades_per_chassis_ + l]
 250               .black_links_[j] = linkdown;
 251           uniqueId++;
 252         }
 253       }
 254     }
 255   }
 256
 257   // Blue links between groups - Not all routers involved, only one per group is linked to others. Let's say router n of
 258   // each group is linked to group n.
 259   // FIXME: in reality blue links may be attached to several different routers
 260   for (unsigned int i = 0; i < num_groups_; i++) {
 261     for (unsigned int j = i + 1; j < num_groups_; j++) {
 262       unsigned int routernumi = i * num_blades_per_chassis_ * num_chassis_per_group_ + j;
 263       unsigned int routernumj = j * num_blades_per_chassis_ * num_chassis_per_group_ + i;
 264       std::string id = "blue_link_between_group_" + std::to_string(i) + "_and_" + std::to_string(j) + "_routers_" +
 265                        std::to_string(routernumi) + "_and_" + std::to_string(routernumj) + "_" +
 266                        std::to_string(uniqueId);
 267       generate_link(id, num_links_blue_, &linkup, &linkdown);
 268
 269       routers_[routernumi].blue_link_ = linkup;
 270       routers_[routernumj].blue_link_ = linkdown;
 271       uniqueId++;
 272     }
 273   }
 274 }
 275
 276 void DragonflyZone::get_local_route(NetPoint* src, NetPoint* dst, Route* route, double* latency)
 277 {
 278   // Minimal routing version.
 279   // TODO : non-minimal random one, and adaptive ?
 280
 281   if (dst->is_router() || src->is_router())
 282     return;
 283
 284   XBT_VERB("dragonfly getLocalRoute from '%s'[%u] to '%s'[%u]", src->get_cname(), src->id(), dst->get_cname(),
 285            dst->id());
 286
 287   if ((src->id() == dst->id()) && has_loopback()) {
 288     resource::LinkImpl* uplink = get_uplink_from(node_pos(src->id()));
 289
 290     route->link_list_.push_back(uplink);
 291     if (latency)
 292       *latency += uplink->get_latency();
 293     return;
 294   }
 295
 296   const auto myCoords     = rankId_to_coords(src->id());
 297   const auto targetCoords = rankId_to_coords(dst->id());
 298   XBT_DEBUG("src : %u group, %u chassis, %u blade, %u node", myCoords.group, myCoords.chassis, myCoords.blade,
 299             myCoords.node);
 300   XBT_DEBUG("dst : %u group, %u chassis, %u blade, %u node", targetCoords.group, targetCoords.chassis,
 301             targetCoords.blade, targetCoords.node);
 302
 303   DragonflyRouter* myRouter      = &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
 304                                         myCoords.chassis * num_blades_per_chassis_ + myCoords.blade];
 305   DragonflyRouter* targetRouter  = &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
 306                                             targetCoords.chassis * num_blades_per_chassis_ + targetCoords.blade];
 307   DragonflyRouter* currentRouter = myRouter;
 308
 309   if (has_limiter()) { // limiter for sender
 310     route->link_list_.push_back(get_uplink_from(node_pos_with_loopback(src->id())));
 311   }
 312
 313   // node->router local link
 314   route->link_list_.push_back(myRouter->my_nodes_[myCoords.node * num_links_per_link_]);
 315   if (latency)
 316     *latency += myRouter->my_nodes_[myCoords.node * num_links_per_link_]->get_latency();
 317
 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         route->link_list_.push_back(currentRouter->green_links_[targetCoords.group]);
 327         if (latency)
 328           *latency += currentRouter->green_links_[targetCoords.group]->get_latency();
 329         currentRouter = &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
 330                                   myCoords.chassis * num_blades_per_chassis_ + targetCoords.group];
 331       }
 332
 333       if (currentRouter->chassis_ != 0) {
 334         // go to the first chassis of our group
 335         if (currentRouter->limiter_)
 336           route->link_list_.push_back(currentRouter->limiter_);
 337         route->link_list_.push_back(currentRouter->black_links_[0]);
 338         if (latency)
 339           *latency += currentRouter->black_links_[0]->get_latency();
 340         currentRouter =
 341             &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + targetCoords.group];
 342       }
 343
 344       // go to destination group - the only optical hop
 345       route->link_list_.push_back(currentRouter->blue_link_);
 346       if (currentRouter->limiter_)
 347         route->link_list_.push_back(currentRouter->limiter_);
 348       if (latency)
 349         *latency += currentRouter->blue_link_->get_latency();
 350       currentRouter =
 351           &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + myCoords.group];
 352     }
 353
 354     // same group, but same blade ?
 355     if (targetRouter->blade_ != currentRouter->blade_) {
 356       if (currentRouter->limiter_)
 357         route->link_list_.push_back(currentRouter->limiter_);
 358       route->link_list_.push_back(currentRouter->green_links_[targetCoords.blade]);
 359       if (latency)
 360         *latency += currentRouter->green_links_[targetCoords.blade]->get_latency();
 361       currentRouter =
 362           &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + targetCoords.blade];
 363     }
 364
 365     // same blade, but same chassis ?
 366     if (targetRouter->chassis_ != currentRouter->chassis_) {
 367       if (currentRouter->limiter_)
 368         route->link_list_.push_back(currentRouter->limiter_);
 369       route->link_list_.push_back(currentRouter->black_links_[targetCoords.chassis]);
 370       if (latency)
 371         *latency += currentRouter->black_links_[targetCoords.chassis]->get_latency();
 372     }
 373   }
 374
 375   // router->node local link
 376   if (targetRouter->limiter_)
 377     route->link_list_.push_back(targetRouter->limiter_);
 378   route->link_list_.push_back(
 379       targetRouter->my_nodes_[targetCoords.node * num_links_per_link_ + num_links_per_link_ - 1]);
 380
 381   if (latency)
 382     *latency +=
 383         targetRouter->my_nodes_[targetCoords.node * num_links_per_link_ + num_links_per_link_ - 1]->get_latency();
 384
 385   if (has_limiter()) { // limiter for receiver
 386     route->link_list_.push_back(get_downlink_to(node_pos_with_loopback(dst->id())));
 387   }
 388
 389   // set gateways (if any)
 390   route->gw_src_ = get_gateway(src->id());
 391   route->gw_dst_ = get_gateway(dst->id());
 392 }
 393 } // namespace routing
 394 } // namespace kernel
 395
 396 namespace s4u {
 397 DragonflyParams::DragonflyParams(const std::pair<unsigned int, unsigned int>& groups,
 398                                  const std::pair<unsigned int, unsigned int>& chassis,
 399                                  const std::pair<unsigned int, unsigned int>& routers, unsigned int nodes)
 400     : groups(groups), chassis(chassis), routers(routers), nodes(nodes)
 401 {
 402   if (groups.first == 0)
 403     throw std::invalid_argument("Dragonfly: Invalid number of groups, must be > 0");
 404   if (groups.second == 0)
 405     throw std::invalid_argument("Dragonfly: Invalid number of blue (groups) links, must be > 0");
 406   if (chassis.first == 0)
 407     throw std::invalid_argument("Dragonfly: Invalid number of chassis, must be > 0");
 408   if (chassis.second == 0)
 409     throw std::invalid_argument("Dragonfly: Invalid number of black (chassis) links, must be > 0");
 410   if (routers.first == 0)
 411     throw std::invalid_argument("Dragonfly: Invalid number of routers, must be > 0");
 412   if (routers.second == 0)
 413     throw std::invalid_argument("Dragonfly: Invalid number of green (routers) links, must be > 0");
 414   if (nodes == 0)
 415     throw std::invalid_argument("Dragonfly: Invalid number of nodes, must be > 0");
 416 }
 417
 418 NetZone* create_dragonfly_zone(const std::string& name, const NetZone* parent, const DragonflyParams& params,
 419                                const ClusterCallbacks& set_callbacks, double bandwidth, double latency,
 420                                Link::SharingPolicy sharing_policy)
 421 {
 422   /* initial checks */
 423   if (bandwidth <= 0)
 424     throw std::invalid_argument("DragonflyZone: incorrect bandwidth for internode communication, bw=" +
 425                                 std::to_string(bandwidth));
 426   if (latency < 0)
 427     throw std::invalid_argument("DragonflyZone: incorrect latency for internode communication, lat=" +
 428                                 std::to_string(latency));
 429
 430   /* creating zone */
 431   auto* zone = new kernel::routing::DragonflyZone(name);
 432   zone->set_topology(params.groups.first, params.groups.second, params.chassis.first, params.chassis.second,
 433                      params.routers.first, params.routers.second, params.nodes);
 434   if (parent)
 435     zone->set_parent(parent->get_impl());
 436   zone->set_link_characteristics(bandwidth, latency, sharing_policy);
 437
 438   /* populating it */
 439   std::vector<unsigned int> dimensions = {params.groups.first, params.chassis.first, params.routers.first,
 440                                           params.nodes};
 441   int tot_elements                     = std::accumulate(dimensions.begin(), dimensions.end(), 1, std::multiplies<>());
 442   for (int i = 0; i < tot_elements; i++) {
 443     kernel::routing::NetPoint* netpoint;
 444     Link* limiter;
 445     Link* loopback;
 446     zone->fill_leaf_from_cb(i, dimensions, set_callbacks, &netpoint, &loopback, &limiter);
 447   }
 448   zone->build_upper_levels(set_callbacks);
 449
 450   return zone->get_iface();
 451 }
 452 } // namespace s4u
 453
 454 } // namespace simgrid