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 #include "src/surf/xml/platf_private.hpp"
  10
  11 #include <boost/algorithm/string/classification.hpp>
  12 #include <boost/algorithm/string/split.hpp>
  13 #include <numeric>
  14 #include <string>
  15
  16 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_route_cluster_dragonfly, surf_route_cluster, "Dragonfly Routing part of surf");
  17
  18 namespace simgrid {
  19 namespace kernel {
  20 namespace routing {
  21
  22 DragonflyZone::DragonflyZone(const std::string& name) : ClusterZone(name) {}
  23
  24 DragonflyZone::Coords DragonflyZone::rankId_to_coords(int rankId) const
  25 {
  26   // coords : group, chassis, blade, node
  27   Coords coords;
  28   coords.group   = rankId / (num_chassis_per_group_ * num_blades_per_chassis_ * num_nodes_per_blade_);
  29   rankId         = rankId % (num_chassis_per_group_ * num_blades_per_chassis_ * num_nodes_per_blade_);
  30   coords.chassis = rankId / (num_blades_per_chassis_ * num_nodes_per_blade_);
  31   rankId         = rankId % (num_blades_per_chassis_ * num_nodes_per_blade_);
  32   coords.blade   = rankId / num_nodes_per_blade_;
  33   coords.node    = rankId % num_nodes_per_blade_;
  34   return coords;
  35 }
  36
  37 void DragonflyZone::rankId_to_coords(int rankId, unsigned int coords[4]) const // XBT_ATTRIB_DEPRECATED_v330
  38 {
  39   const auto s_coords = rankId_to_coords(rankId);
  40   coords[0]           = s_coords.group;
  41   coords[1]           = s_coords.chassis;
  42   coords[2]           = s_coords.blade;
  43   coords[3]           = s_coords.node;
  44 }
  45
  46 void DragonflyZone::set_link_characteristics(double bw, double lat, s4u::Link::SharingPolicy sharing_policy)
  47 {
  48   sharing_policy_ = sharing_policy;
  49   if (sharing_policy == s4u::Link::SharingPolicy::SPLITDUPLEX)
  50     num_links_per_link_ = 2;
  51   bw_  = bw;
  52   lat_ = lat;
  53 }
  54
  55 void DragonflyZone::set_topology(unsigned int n_groups, unsigned int groups_links, unsigned int n_chassis,
  56                                  unsigned int chassis_links, unsigned int n_routers, unsigned int routers_links,
  57                                  unsigned int nodes)
  58 {
  59   num_groups_     = n_groups;
  60   num_links_blue_ = groups_links;
  61
  62   num_chassis_per_group_ = n_chassis;
  63   num_links_black_       = chassis_links;
  64
  65   num_blades_per_chassis_ = n_routers;
  66   num_links_green_        = routers_links;
  67
  68   num_nodes_per_blade_ = nodes;
  69 }
  70
  71 void DragonflyZone::parse_specific_arguments(ClusterCreationArgs* cluster)
  72 {
  73   std::vector<std::string> parameters;
  74   std::vector<std::string> tmp;
  75   boost::split(parameters, cluster->topo_parameters, boost::is_any_of(";"));
  76
  77   if (parameters.size() != 4)
  78     surf_parse_error(
  79         "Dragonfly are defined by the number of groups, chassis per groups, blades per chassis, nodes per blade");
  80
  81   // Blue network : number of groups, number of links between each group
  82   boost::split(tmp, parameters[0], boost::is_any_of(","));
  83   if (tmp.size() != 2)
  84     surf_parse_error("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
  85
  86   try {
  87     num_groups_ = std::stoi(tmp[0]);
  88   } catch (const std::invalid_argument&) {
  89     throw std::invalid_argument(std::string("Invalid number of groups:") + tmp[0]);
  90   }
  91
  92   try {
  93     num_links_blue_ = std::stoi(tmp[1]);
  94   } catch (const std::invalid_argument&) {
  95     throw std::invalid_argument(std::string("Invalid number of links for the blue level:") + tmp[1]);
  96   }
  97
  98   // Black network : number of chassis/group, number of links between each router on the black network
  99   boost::split(tmp, parameters[1], boost::is_any_of(","));
 100   if (tmp.size() != 2)
 101     surf_parse_error("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
 102
 103   try {
 104     num_chassis_per_group_ = std::stoi(tmp[0]);
 105   } catch (const std::invalid_argument&) {
 106     throw std::invalid_argument(std::string("Invalid number of groups:") + tmp[0]);
 107   }
 108
 109   try {
 110     num_links_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     surf_parse_error("Dragonfly topologies are defined by 3 levels with 2 elements each, and one with one element");
 119
 120   try {
 121     num_blades_per_chassis_ = std::stoi(tmp[0]);
 122   } catch (const std::invalid_argument&) {
 123     throw std::invalid_argument(std::string("Invalid number of groups:") + tmp[0]);
 124   }
 125
 126   try {
 127     num_links_green_ = std::stoi(tmp[1]);
 128   } catch (const std::invalid_argument&) {
 129     throw std::invalid_argument(std::string("Invalid number of links for the green level:") + tmp[1]);
 130   }
 131
 132   // The last part of topo_parameters should be the number of nodes per blade
 133   try {
 134     num_nodes_per_blade_ = std::stoi(parameters[3]);
 135   } catch (const std::invalid_argument&) {
 136     throw std::invalid_argument(std::string("Last parameter is not the amount of nodes per blade:") + parameters[3]);
 137   }
 138
 139   set_link_characteristics(cluster->bw, cluster->lat, cluster->sharing_policy);
 140 }
 141
 142 /* Generate the cluster once every node is created */
 143 void DragonflyZone::do_seal()
 144 {
 145   if (num_nodes_per_blade_ == 0)
 146     return;
 147
 148   generate_routers();
 149   generate_links();
 150 }
 151
 152 void DragonflyZone::generate_routers()
 153 {
 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         routers_.emplace_back(i, j, k);
 159 }
 160
 161 void DragonflyZone::generate_link(const std::string& id, int numlinks, resource::LinkImpl** linkup,
 162                                   resource::LinkImpl** linkdown)
 163 {
 164   XBT_DEBUG("Generating link %s", id.c_str());
 165   *linkup   = nullptr;
 166   *linkdown = nullptr;
 167   if (sharing_policy_ == s4u::Link::SharingPolicy::SPLITDUPLEX) {
 168     *linkup   = create_link(id + "_UP", std::vector<double>{bw_ * numlinks})->set_latency(lat_)->seal()->get_impl();
 169     *linkdown = create_link(id + "_DOWN", std::vector<double>{bw_ * numlinks})->set_latency(lat_)->seal()->get_impl();
 170   } else {
 171     *linkup   = create_link(id, std::vector<double>{bw_ * numlinks})->set_latency(lat_)->seal()->get_impl();
 172     *linkdown = *linkup;
 173   }
 174 }
 175
 176 void DragonflyZone::generate_links()
 177 {
 178   static int uniqueId = 0;
 179   resource::LinkImpl* linkup;
 180   resource::LinkImpl* linkdown;
 181
 182   unsigned int numRouters = num_groups_ * num_chassis_per_group_ * num_blades_per_chassis_;
 183
 184   // Links from routers to their local nodes.
 185   for (unsigned int i = 0; i < numRouters; i++) {
 186     // allocate structures
 187     routers_[i].my_nodes_.resize(num_links_per_link_ * num_nodes_per_blade_);
 188     routers_[i].green_links_.resize(num_blades_per_chassis_);
 189     routers_[i].black_links_.resize(num_chassis_per_group_);
 190
 191     for (unsigned int j = 0; j < num_links_per_link_ * num_nodes_per_blade_; j += num_links_per_link_) {
 192       std::string id = "local_link_from_router_" + std::to_string(i) + "_to_node_" +
 193                        std::to_string(j / num_links_per_link_) + "_" + std::to_string(uniqueId);
 194       generate_link(id, 1, &linkup, &linkdown);
 195
 196       routers_[i].my_nodes_[j] = linkup;
 197       if (sharing_policy_ == s4u::Link::SharingPolicy::SPLITDUPLEX)
 198         routers_[i].my_nodes_[j + 1] = linkdown;
 199
 200       uniqueId++;
 201     }
 202   }
 203
 204   // Green links from routers to same chassis routers - alltoall
 205   for (unsigned int i = 0; i < num_groups_ * num_chassis_per_group_; i++) {
 206     for (unsigned int j = 0; j < num_blades_per_chassis_; j++) {
 207       for (unsigned int k = j + 1; k < num_blades_per_chassis_; k++) {
 208         std::string id = "green_link_in_chassis_" + std::to_string(i % num_chassis_per_group_) + "_between_routers_" +
 209                          std::to_string(j) + "_and_" + std::to_string(k) + "_" + std::to_string(uniqueId);
 210         generate_link(id, num_links_green_, &linkup, &linkdown);
 211
 212         routers_[i * num_blades_per_chassis_ + j].green_links_[k] = linkup;
 213         routers_[i * num_blades_per_chassis_ + k].green_links_[j] = linkdown;
 214         uniqueId++;
 215       }
 216     }
 217   }
 218
 219   // Black links from routers to same group routers - alltoall
 220   for (unsigned int i = 0; i < num_groups_; i++) {
 221     for (unsigned int j = 0; j < num_chassis_per_group_; j++) {
 222       for (unsigned int k = j + 1; k < num_chassis_per_group_; k++) {
 223         for (unsigned int l = 0; l < num_blades_per_chassis_; l++) {
 224           std::string id = "black_link_in_group_" + std::to_string(i) + "_between_chassis_" + std::to_string(j) +
 225                            "_and_" + std::to_string(k) + "_blade_" + std::to_string(l) + "_" + std::to_string(uniqueId);
 226           generate_link(id, num_links_black_, &linkup, &linkdown);
 227
 228           routers_[i * num_blades_per_chassis_ * num_chassis_per_group_ + j * num_blades_per_chassis_ + l]
 229               .black_links_[k] = linkup;
 230           routers_[i * num_blades_per_chassis_ * num_chassis_per_group_ + k * num_blades_per_chassis_ + l]
 231               .black_links_[j] = linkdown;
 232           uniqueId++;
 233         }
 234       }
 235     }
 236   }
 237
 238   // Blue links between groups - Not all routers involved, only one per group is linked to others. Let's say router n of
 239   // each group is linked to group n.
 240   // FIXME: in reality blue links may be attached to several different routers
 241   for (unsigned int i = 0; i < num_groups_; i++) {
 242     for (unsigned int j = i + 1; j < num_groups_; j++) {
 243       unsigned int routernumi = i * num_blades_per_chassis_ * num_chassis_per_group_ + j;
 244       unsigned int routernumj = j * num_blades_per_chassis_ * num_chassis_per_group_ + i;
 245       std::string id = "blue_link_between_group_" + std::to_string(i) + "_and_" + std::to_string(j) + "_routers_" +
 246                        std::to_string(routernumi) + "_and_" + std::to_string(routernumj) + "_" +
 247                        std::to_string(uniqueId);
 248       generate_link(id, num_links_blue_, &linkup, &linkdown);
 249
 250       routers_[routernumi].blue_link_ = linkup;
 251       routers_[routernumj].blue_link_ = linkdown;
 252       uniqueId++;
 253     }
 254   }
 255 }
 256
 257 void DragonflyZone::get_local_route(NetPoint* src, NetPoint* dst, RouteCreationArgs* route, double* latency)
 258 {
 259   // Minimal routing version.
 260   // TODO : non-minimal random one, and adaptive ?
 261
 262   if (dst->is_router() || src->is_router())
 263     return;
 264
 265   XBT_VERB("dragonfly getLocalRoute from '%s'[%u] to '%s'[%u]", src->get_cname(), src->id(), dst->get_cname(),
 266            dst->id());
 267
 268   if ((src->id() == dst->id()) && has_loopback()) {
 269     resource::LinkImpl* uplink = get_uplink_from(node_pos(src->id()));
 270
 271     route->link_list.push_back(uplink);
 272     if (latency)
 273       *latency += uplink->get_latency();
 274     return;
 275   }
 276
 277   const auto myCoords     = rankId_to_coords(src->id());
 278   const auto targetCoords = rankId_to_coords(dst->id());
 279   XBT_DEBUG("src : %u group, %u chassis, %u blade, %u node", myCoords.group, myCoords.chassis, myCoords.blade,
 280             myCoords.node);
 281   XBT_DEBUG("dst : %u group, %u chassis, %u blade, %u node", targetCoords.group, targetCoords.chassis,
 282             targetCoords.blade, targetCoords.node);
 283
 284   DragonflyRouter* myRouter      = &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
 285                                         myCoords.chassis * num_blades_per_chassis_ + myCoords.blade];
 286   DragonflyRouter* targetRouter  = &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
 287                                             targetCoords.chassis * num_blades_per_chassis_ + targetCoords.blade];
 288   DragonflyRouter* currentRouter = myRouter;
 289
 290   // node->router local link
 291   route->link_list.push_back(myRouter->my_nodes_[myCoords.node * num_links_per_link_]);
 292   if (latency)
 293     *latency += myRouter->my_nodes_[myCoords.node * num_links_per_link_]->get_latency();
 294
 295   if (has_limiter()) { // limiter for sender
 296     route->link_list.push_back(get_uplink_from(node_pos_with_loopback(src->id())));
 297   }
 298
 299   if (targetRouter != myRouter) {
 300     // are we on a different group ?
 301     if (targetRouter->group_ != currentRouter->group_) {
 302       // go to the router of our group connected to this one.
 303       if (currentRouter->blade_ != targetCoords.group) {
 304         // go to the nth router in our chassis
 305         route->link_list.push_back(currentRouter->green_links_[targetCoords.group]);
 306         if (latency)
 307           *latency += currentRouter->green_links_[targetCoords.group]->get_latency();
 308         currentRouter = &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) +
 309                                   myCoords.chassis * num_blades_per_chassis_ + targetCoords.group];
 310       }
 311
 312       if (currentRouter->chassis_ != 0) {
 313         // go to the first chassis of our group
 314         route->link_list.push_back(currentRouter->black_links_[0]);
 315         if (latency)
 316           *latency += currentRouter->black_links_[0]->get_latency();
 317         currentRouter =
 318             &routers_[myCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + targetCoords.group];
 319       }
 320
 321       // go to destination group - the only optical hop
 322       route->link_list.push_back(currentRouter->blue_link_);
 323       if (latency)
 324         *latency += currentRouter->blue_link_->get_latency();
 325       currentRouter =
 326           &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + myCoords.group];
 327     }
 328
 329     // same group, but same blade ?
 330     if (targetRouter->blade_ != currentRouter->blade_) {
 331       route->link_list.push_back(currentRouter->green_links_[targetCoords.blade]);
 332       if (latency)
 333         *latency += currentRouter->green_links_[targetCoords.blade]->get_latency();
 334       currentRouter =
 335           &routers_[targetCoords.group * (num_chassis_per_group_ * num_blades_per_chassis_) + targetCoords.blade];
 336     }
 337
 338     // same blade, but same chassis ?
 339     if (targetRouter->chassis_ != currentRouter->chassis_) {
 340       route->link_list.push_back(currentRouter->black_links_[targetCoords.chassis]);
 341       if (latency)
 342         *latency += currentRouter->black_links_[targetCoords.chassis]->get_latency();
 343     }
 344   }
 345
 346   if (has_limiter()) { // limiter for receiver
 347     route->link_list.push_back(get_downlink_to(node_pos_with_loopback(dst->id())));
 348   }
 349
 350   // router->node local link
 351   route->link_list.push_back(
 352       targetRouter->my_nodes_[targetCoords.node * num_links_per_link_ + num_links_per_link_ - 1]);
 353   if (latency)
 354     *latency +=
 355         targetRouter->my_nodes_[targetCoords.node * num_links_per_link_ + num_links_per_link_ - 1]->get_latency();
 356
 357   // set gateways (if any)
 358   route->gw_src = get_gateway(src->id());
 359   route->gw_dst = get_gateway(dst->id());
 360 }
 361 } // namespace routing
 362 } // namespace kernel
 363
 364 namespace s4u {
 365 DragonflyParams::DragonflyParams(const std::pair<unsigned int, unsigned int>& groups,
 366                                  const std::pair<unsigned int, unsigned int>& chassis,
 367                                  const std::pair<unsigned int, unsigned int>& routers, unsigned int nodes)
 368     : groups(groups), chassis(chassis), routers(routers), nodes(nodes)
 369 {
 370   if (groups.first == 0)
 371     throw std::invalid_argument("Dragonfly: Invalid number of groups, must be > 0");
 372   if (groups.second == 0)
 373     throw std::invalid_argument("Dragonfly: Invalid number of blue (groups) links, must be > 0");
 374   if (chassis.first == 0)
 375     throw std::invalid_argument("Dragonfly: Invalid number of chassis, must be > 0");
 376   if (chassis.second == 0)
 377     throw std::invalid_argument("Dragonfly: Invalid number of black (chassis) links, must be > 0");
 378   if (routers.first == 0)
 379     throw std::invalid_argument("Dragonfly: Invalid number of routers, must be > 0");
 380   if (routers.second == 0)
 381     throw std::invalid_argument("Dragonfly: Invalid number of green (routers) links, must be > 0");
 382   if (nodes == 0)
 383     throw std::invalid_argument("Dragonfly: Invalid number of nodes, must be > 0");
 384 }
 385
 386 NetZone* create_dragonfly_zone(const std::string& name, const NetZone* parent, const DragonflyParams& params,
 387                                double bandwidth, double latency, Link::SharingPolicy sharing_policy,
 388                                const std::function<ClusterNetPointCb>& set_netpoint,
 389                                const std::function<ClusterLinkCb>& set_loopback,
 390                                const std::function<ClusterLinkCb>& set_limiter)
 391 {
 392   /* initial checks */
 393   if (bandwidth <= 0)
 394     throw std::invalid_argument("DragonflyZone: incorrect bandwidth for internode communication, bw=" +
 395                                 std::to_string(bandwidth));
 396   if (latency < 0)
 397     throw std::invalid_argument("DragonflyZone: incorrect latency for internode communication, lat=" +
 398                                 std::to_string(latency));
 399
 400   /* creating zone */
 401   auto* zone = new kernel::routing::DragonflyZone(name);
 402   zone->set_topology(params.groups.first, params.groups.second, params.chassis.first, params.chassis.second,
 403                      params.routers.first, params.routers.second, params.nodes);
 404   if (parent)
 405     zone->set_parent(parent->get_impl());
 406   zone->set_link_characteristics(bandwidth, latency, sharing_policy);
 407
 408   /* populating it */
 409   std::vector<unsigned int> dimensions = {params.groups.first, params.chassis.first, params.routers.first,
 410                                           params.nodes};
 411   int tot_elements                     = std::accumulate(dimensions.begin(), dimensions.end(), 1, std::multiplies<>());
 412   for (int i = 0; i < tot_elements; i++) {
 413     kernel::routing::NetPoint* netpoint;
 414     Link* limiter;
 415     Link* loopback;
 416     zone->fill_leaf_from_cb(i, dimensions, set_netpoint, set_loopback, set_limiter, &netpoint, &loopback, &limiter);
 417   }
 418
 419   return zone->get_iface();
 420 }
 421 } // namespace s4u
 422
 423 } // namespace simgrid