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) : ClusterBase(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   ClusterBase::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       id--;
 160       const 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     }
 166     return limiter;
 167   };
 168
 169   routers_.reserve(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));
 174       }
 175     }
 176   }
 177 }
 178
 179 void DragonflyZone::generate_link(const std::string& id, int numlinks, resource::LinkImpl** linkup,
 180                                   resource::LinkImpl** linkdown)
 181 {
 182   XBT_DEBUG("Generating link %s", id.c_str());
 183   *linkup   = nullptr;
 184   *linkdown = nullptr;
 185   if (get_link_sharing_policy() == s4u::Link::SharingPolicy::SPLITDUPLEX) {
 186     *linkup = create_link(id + "_UP", std::vector<double>{get_link_bandwidth() * numlinks})
 187                   ->set_latency(get_link_latency())
 188                   ->seal()
 189                   ->get_impl();
 190     *linkdown = create_link(id + "_DOWN", std::vector<double>{get_link_bandwidth() * numlinks})
 191                     ->set_latency(get_link_latency())
 192                     ->seal()
 193                     ->get_impl();
 194   } else {
 195     *linkup = create_link(id, std::vector<double>{get_link_bandwidth() * numlinks})
 196                   ->set_latency(get_link_latency())
 197                   ->seal()
 198                   ->get_impl();
 199     *linkdown = *linkup;
 200   }
 201 }
 202
 203 void DragonflyZone::generate_links()
 204 {
 205   static int uniqueId = 0;
 206   resource::LinkImpl* linkup;
 207   resource::LinkImpl* linkdown;
 208
 209   unsigned int numRouters = num_groups_ * num_chassis_per_group_ * num_blades_per_chassis_;
 210
 211   // Links from routers to their local nodes.
 212   for (unsigned int i = 0; i < numRouters; i++) {
 213     // allocate structures
 214     routers_[i].my_nodes_.resize(num_links_per_link_ * num_nodes_per_blade_);
 215     routers_[i].green_links_.resize(num_blades_per_chassis_);
 216     routers_[i].black_links_.resize(num_chassis_per_group_);
 217
 218     for (unsigned int j = 0; j < num_links_per_link_ * num_nodes_per_blade_; j += num_links_per_link_) {
 219       std::string id = "local_link_from_router_" + std::to_string(i) + "_to_node_" +
 220                        std::to_string(j / num_links_per_link_) + "_" + std::to_string(uniqueId);
 221       generate_link(id, 1, &linkup, &linkdown);
 222
 223       routers_[i].my_nodes_[j] = linkup;
 224       if (get_link_sharing_policy() == s4u::Link::SharingPolicy::SPLITDUPLEX)
 225         routers_[i].my_nodes_[j + 1] = linkdown;
 226
 227       uniqueId++;
 228     }
 229   }
 230
 231   // Green links from routers to same chassis routers - alltoall
 232   for (unsigned int i = 0; i < num_groups_ * num_chassis_per_group_; i++) {
 233     for (unsigned int j = 0; j < num_blades_per_chassis_; j++) {
 234       for (unsigned int k = j + 1; k < num_blades_per_chassis_; k++) {
 235         std::string id = "green_link_in_chassis_" + std::to_string(i % num_chassis_per_group_) + "_between_routers_" +
 236                          std::to_string(j) + "_and_" + std::to_string(k) + "_" + std::to_string(uniqueId);
 237         generate_link(id, num_links_green_, &linkup, &linkdown);
 238
 239         routers_[i * num_blades_per_chassis_ + j].green_links_[k] = linkup;
 240         routers_[i * num_blades_per_chassis_ + k].green_links_[j] = linkdown;
 241         uniqueId++;
 242       }
 243     }
 244   }
 245
 246   // Black links from routers to same group routers - alltoall
 247   for (unsigned int i = 0; i < num_groups_; i++) {
 248     for (unsigned int j = 0; j < num_chassis_per_group_; j++) {
 249       for (unsigned int k = j + 1; k < num_chassis_per_group_; k++) {
 250         for (unsigned int l = 0; l < num_blades_per_chassis_; l++) {
 251           std::string id = "black_link_in_group_" + std::to_string(i) + "_between_chassis_" + std::to_string(j) +
 252                            "_and_" + std::to_string(k) + "_blade_" + std::to_string(l) + "_" + std::to_string(uniqueId);
 253           generate_link(id, num_links_black_, &linkup, &linkdown);
 254
 255           routers_[i * num_blades_per_chassis_ * num_chassis_per_group_ + j * num_blades_per_chassis_ + l]
 256               .black_links_[k] = linkup;
 257           routers_[i * num_blades_per_chassis_ * num_chassis_per_group_ + k * num_blades_per_chassis_ + l]
 258               .black_links_[j] = linkdown;
 259           uniqueId++;
 260         }
 261       }
 262     }
 263   }
 264
 265   // Blue links between groups - Not all routers involved, only one per group is linked to others. Let's say router n of
 266   // each group is linked to group n.
 267   // FIXME: in reality blue links may be attached to several different routers
 268   for (unsigned int i = 0; i < num_groups_; i++) {
 269     for (unsigned int j = i + 1; j < num_groups_; j++) {
 270       unsigned int routernumi = i * num_blades_per_chassis_ * num_chassis_per_group_ + j;
 271       unsigned int routernumj = j * num_blades_per_chassis_ * num_chassis_per_group_ + i;
 272       std::string id = "blue_link_between_group_" + std::to_string(i) + "_and_" + std::to_string(j) + "_routers_" +
 273                        std::to_string(routernumi) + "_and_" + std::to_string(routernumj) + "_" +
 274                        std::to_string(uniqueId);
 275       generate_link(id, num_links_blue_, &linkup, &linkdown);
 276
 277       routers_[routernumi].blue_link_ = linkup;
 278       routers_[routernumj].blue_link_ = linkdown;
 279       uniqueId++;
 280     }
 281   }
 282 }
 283
 284 void DragonflyZone::get_local_route(const NetPoint* src, const NetPoint* dst, Route* route, double* latency)
 285 {
 286   // Minimal routing version.
 287   // TODO : non-minimal random one, and adaptive ?
 288
 289   if (dst->is_router() || src->is_router())
 290     return;
 291
 292   XBT_VERB("dragonfly getLocalRoute from '%s'[%u] to '%s'[%u]", src->get_cname(), src->id(), dst->get_cname(),
 293            dst->id());
 294
 295   if ((src->id() == dst->id()) && has_loopback()) {
 296     resource::LinkImpl* uplink = get_uplink_from(node_pos(src->id()));
 297
 298     route->link_list_.push_back(uplink);
 299     if (latency)
 300       *latency += uplink->get_latency();
 301     return;
 302   }
 303
 304   const auto myCoords     = rankId_to_coords(src->id());
 305   const auto targetCoords = rankId_to_coords(dst->id());
 306   XBT_DEBUG("src : %u group, %u chassis, %u blade, %u node", myCoords.group, myCoords.chassis, myCoords.blade,
 307             myCoords.node);
 308   XBT_DEBUG("dst : %u group, %u chassis, %u blade, %u node", targetCoords.group, targetCoords.chassis,
 309             targetCoords.blade, targetCoords.node);
 310
 311   DragonflyRouter* myRouter      = &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
 312                                         myCoords.chassis * num_blades_per_chassis_ + myCoords.blade];
 313   DragonflyRouter* targetRouter  = &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
 314                                             targetCoords.chassis * num_blades_per_chassis_ + targetCoords.blade];
 315   DragonflyRouter* currentRouter = myRouter;
 316
 317   if (has_limiter()) { // limiter for sender
 318     route->link_list_.push_back(get_uplink_from(node_pos_with_loopback(src->id())));
 319   }
 320
 321   // node->router local link
 322   route->link_list_.push_back(myRouter->my_nodes_[myCoords.node * num_links_per_link_]);
 323   if (latency)
 324     *latency += myRouter->my_nodes_[myCoords.node * num_links_per_link_]->get_latency();
 325
 326   if (targetRouter != myRouter) {
 327     // are we on a different group ?
 328     if (targetRouter->group_ != currentRouter->group_) {
 329       // go to the router of our group connected to this one.
 330       if (currentRouter->blade_ != targetCoords.group) {
 331         if (currentRouter->limiter_)
 332           route->link_list_.push_back(currentRouter->limiter_);
 333         // go to the nth router in our chassis
 334         route->link_list_.push_back(currentRouter->green_links_[targetCoords.group]);
 335         if (latency)
 336           *latency += currentRouter->green_links_[targetCoords.group]->get_latency();
 337         currentRouter = &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
 338                                   myCoords.chassis * num_blades_per_chassis_ + targetCoords.group];
 339       }
 340
 341       if (currentRouter->chassis_ != 0) {
 342         // go to the first chassis of our group
 343         if (currentRouter->limiter_)
 344           route->link_list_.push_back(currentRouter->limiter_);
 345         route->link_list_.push_back(currentRouter->black_links_[0]);
 346         if (latency)
 347           *latency += currentRouter->black_links_[0]->get_latency();
 348         currentRouter =
 349             &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + targetCoords.group];
 350       }
 351
 352       // go to destination group - the only optical hop
 353       route->link_list_.push_back(currentRouter->blue_link_);
 354       if (currentRouter->limiter_)
 355         route->link_list_.push_back(currentRouter->limiter_);
 356       if (latency)
 357         *latency += currentRouter->blue_link_->get_latency();
 358       currentRouter =
 359           &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + myCoords.group];
 360     }
 361
 362     // same group, but same blade ?
 363     if (targetRouter->blade_ != currentRouter->blade_) {
 364       if (currentRouter->limiter_)
 365         route->link_list_.push_back(currentRouter->limiter_);
 366       route->link_list_.push_back(currentRouter->green_links_[targetCoords.blade]);
 367       if (latency)
 368         *latency += currentRouter->green_links_[targetCoords.blade]->get_latency();
 369       currentRouter =
 370           &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + targetCoords.blade];
 371     }
 372
 373     // same blade, but same chassis ?
 374     if (targetRouter->chassis_ != currentRouter->chassis_) {
 375       if (currentRouter->limiter_)
 376         route->link_list_.push_back(currentRouter->limiter_);
 377       route->link_list_.push_back(currentRouter->black_links_[targetCoords.chassis]);
 378       if (latency)
 379         *latency += currentRouter->black_links_[targetCoords.chassis]->get_latency();
 380     }
 381   }
 382
 383   // router->node local link
 384   if (targetRouter->limiter_)
 385     route->link_list_.push_back(targetRouter->limiter_);
 386   route->link_list_.push_back(
 387       targetRouter->my_nodes_[targetCoords.node * num_links_per_link_ + num_links_per_link_ - 1]);
 388
 389   if (latency)
 390     *latency +=
 391         targetRouter->my_nodes_[targetCoords.node * num_links_per_link_ + num_links_per_link_ - 1]->get_latency();
 392
 393   if (has_limiter()) { // limiter for receiver
 394     route->link_list_.push_back(get_downlink_to(node_pos_with_loopback(dst->id())));
 395   }
 396
 397   // set gateways (if any)
 398   route->gw_src_ = get_gateway(src->id());
 399   route->gw_dst_ = get_gateway(dst->id());
 400 }
 401 } // namespace routing
 402 } // namespace kernel
 403
 404 namespace s4u {
 405 DragonflyParams::DragonflyParams(const std::pair<unsigned int, unsigned int>& groups,
 406                                  const std::pair<unsigned int, unsigned int>& chassis,
 407                                  const std::pair<unsigned int, unsigned int>& routers, unsigned int nodes)
 408     : groups(groups), chassis(chassis), routers(routers), nodes(nodes)
 409 {
 410   if (groups.first == 0)
 411     throw std::invalid_argument("Dragonfly: Invalid number of groups, must be > 0");
 412   if (groups.second == 0)
 413     throw std::invalid_argument("Dragonfly: Invalid number of blue (groups) links, must be > 0");
 414   if (chassis.first == 0)
 415     throw std::invalid_argument("Dragonfly: Invalid number of chassis, must be > 0");
 416   if (chassis.second == 0)
 417     throw std::invalid_argument("Dragonfly: Invalid number of black (chassis) links, must be > 0");
 418   if (routers.first == 0)
 419     throw std::invalid_argument("Dragonfly: Invalid number of routers, must be > 0");
 420   if (routers.second == 0)
 421     throw std::invalid_argument("Dragonfly: Invalid number of green (routers) links, must be > 0");
 422   if (nodes == 0)
 423     throw std::invalid_argument("Dragonfly: Invalid number of nodes, must be > 0");
 424 }
 425
 426 NetZone* create_dragonfly_zone(const std::string& name, const NetZone* parent, const DragonflyParams& params,
 427                                const ClusterCallbacks& set_callbacks, double bandwidth, double latency,
 428                                Link::SharingPolicy sharing_policy)
 429 {
 430   /* initial checks */
 431   if (bandwidth <= 0)
 432     throw std::invalid_argument("DragonflyZone: incorrect bandwidth for internode communication, bw=" +
 433                                 std::to_string(bandwidth));
 434   if (latency < 0)
 435     throw std::invalid_argument("DragonflyZone: incorrect latency for internode communication, lat=" +
 436                                 std::to_string(latency));
 437
 438   /* creating zone */
 439   auto* zone = new kernel::routing::DragonflyZone(name);
 440   zone->set_topology(params.groups.first, params.groups.second, params.chassis.first, params.chassis.second,
 441                      params.routers.first, params.routers.second, params.nodes);
 442   if (parent)
 443     zone->set_parent(parent->get_impl());
 444   zone->set_link_characteristics(bandwidth, latency, sharing_policy);
 445
 446   /* populating it */
 447   std::vector<unsigned int> dimensions = {params.groups.first, params.chassis.first, params.routers.first,
 448                                           params.nodes};
 449   int tot_elements                     = std::accumulate(dimensions.begin(), dimensions.end(), 1, std::multiplies<>());
 450   for (int i = 0; i < tot_elements; i++) {
 451     kernel::routing::NetPoint* netpoint;
 452     Link* limiter;
 453     Link* loopback;
 454     zone->fill_leaf_from_cb(i, dimensions, set_callbacks, &netpoint, &loopback, &limiter);
 455   }
 456   zone->build_upper_levels(set_callbacks);
 457
 458   return zone->get_iface();
 459 }
 460 } // namespace s4u
 461
 462 } // namespace simgrid