1 /* Copyright (c) 2014-2023. 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/TorusZone.hpp"
7 #include "simgrid/kernel/routing/NetPoint.hpp"
8 #include "simgrid/s4u/Host.hpp"
9 #include "src/kernel/resource/NetworkModel.hpp"
11 #include <boost/algorithm/string/classification.hpp>
12 #include <boost/algorithm/string/split.hpp>
17 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(ker_routing_torus, ker_platform, "Kernel Torus Routing");
20 namespace kernel::routing {
22 void TorusZone::create_torus_links(unsigned long id, int rank, unsigned long position)
24 /* Create all links that exist in the torus. Each rank creates @a dimensions-1 links */
25 unsigned long dim_product = 1; // Needed to calculate the next neighbor_id
27 for (unsigned long j = 0; j < dimensions_.size(); j++) {
28 unsigned long current_dimension =
29 dimensions_[j]; // which dimension are we currently in?
30 // we need to iterate over all dimensions and create all links there
31 // The other node the link connects
32 unsigned long neighbor_rank_id = ((rank / dim_product) % current_dimension == current_dimension - 1)
33 ? rank - (current_dimension - 1) * dim_product
35 // name of neighbor is not right for non contiguous cluster radicals (as id != rank in this case)
36 std::string link_id = get_name() + "_link_from_" + std::to_string(id) + "_to_" + std::to_string(neighbor_rank_id);
37 const s4u::Link* linkup;
38 const s4u::Link* linkdown;
39 if (get_link_sharing_policy() == s4u::Link::SharingPolicy::SPLITDUPLEX) {
40 linkup = create_link(link_id + "_UP", {get_link_bandwidth()})->set_latency(get_link_latency())->seal();
41 linkdown = create_link(link_id + "_DOWN", {get_link_bandwidth()})->set_latency(get_link_latency())->seal();
44 linkup = create_link(link_id, {get_link_bandwidth()})->set_latency(get_link_latency())->seal();
48 * Add the link to its appropriate position.
49 * Note that position rankId*(xbt_dynar_length(dimensions)+has_loopback?+has_limiter?)
50 * holds the link "rankId->rankId"
52 add_private_link_at(position + j, {linkup->get_impl(), linkdown->get_impl()});
53 dim_product *= current_dimension;
57 std::vector<unsigned long> TorusZone::parse_topo_parameters(const std::string& topo_parameters)
59 std::vector<std::string> dimensions_str;
60 boost::split(dimensions_str, topo_parameters, boost::is_any_of(","));
61 std::vector<unsigned long> dimensions;
63 /* We are in a torus cluster
64 * Parse attribute dimensions="dim1,dim2,dim3,...,dimN" and save them into a vector.
65 * Additionally, we need to know how many ranks we have in total
67 std::transform(begin(dimensions_str), end(dimensions_str), std::back_inserter(dimensions),
68 [](const std::string& s) { return std::stoi(s); });
73 void TorusZone::set_topology(const std::vector<unsigned long>& dimensions)
75 xbt_assert(not dimensions.empty(), "Torus dimensions cannot be empty");
76 dimensions_ = dimensions;
77 set_num_links_per_node(dimensions_.size());
80 void TorusZone::get_local_route(const NetPoint* src, const NetPoint* dst, Route* route, double* lat)
82 XBT_VERB("torus getLocalRoute from '%s'[%lu] to '%s'[%lu]", src->get_cname(), src->id(), dst->get_cname(), dst->id());
84 if (dst->is_router() || src->is_router())
87 if (src->id() == dst->id() && has_loopback()) {
88 resource::StandardLinkImpl* uplink = get_uplink_from(node_pos(src->id()));
90 add_link_latency(route->link_list_, uplink, lat);
95 * Dimension based routing routes through each dimension consecutively
96 * TODO Change to dynamic assignment
100 * Arrays that hold the coordinates of the current node and the target; comparing the values at the i-th position of
101 * both arrays, we can easily assess whether we need to route into this dimension or not.
103 const unsigned long dsize = dimensions_.size();
104 std::vector<unsigned long> myCoords(dsize);
105 std::vector<unsigned long> targetCoords(dsize);
106 unsigned int dim_size_product = 1;
107 for (unsigned long i = 0; i < dsize; i++) {
108 unsigned long cur_dim_size = dimensions_[i];
109 myCoords[i] = (src->id() / dim_size_product) % cur_dim_size;
110 targetCoords[i] = (dst->id() / dim_size_product) % cur_dim_size;
111 dim_size_product *= cur_dim_size;
115 * linkOffset describes the offset where the link we want to use is stored(+1 is added because each node has a link
116 * from itself to itself, which can only be the case if src->m_id == dst->m_id -- see above for this special case)
118 unsigned long linkOffset = (dsize + 1) * src->id();
120 bool use_lnk_up = false; // Is this link of the form "cur -> next" or "next -> cur"? false means: next -> cur
121 unsigned long current_node = src->id();
122 while (current_node != dst->id()) {
123 unsigned long next_node = 0;
124 unsigned long dim_product = 1; // First, we will route in x-dimension
125 for (unsigned long j = 0; j < dsize; j++) {
126 const unsigned long cur_dim = dimensions_[j];
127 // current_node/dim_product = position in current dimension
128 if ((current_node / dim_product) % cur_dim != (dst->id() / dim_product) % cur_dim) {
129 if ((targetCoords[j] > myCoords[j] &&
130 targetCoords[j] <= myCoords[j] + cur_dim / 2) // Is the target node on the right, without the wrap-around?
132 (myCoords[j] > cur_dim / 2 && (myCoords[j] + cur_dim / 2) % cur_dim >=
133 targetCoords[j])) { // Or do we need to use the wrap around to reach it?
134 if ((current_node / dim_product) % cur_dim == cur_dim - 1)
135 next_node = (current_node + dim_product - dim_product * cur_dim);
137 next_node = (current_node + dim_product);
139 // HERE: We use *CURRENT* node for calculation (as opposed to next_node)
140 linkOffset = node_pos_with_loopback_limiter(current_node) + j;
142 } else { // Route to the left
143 if ((current_node / dim_product) % cur_dim == 0)
144 next_node = (current_node - dim_product + dim_product * cur_dim);
146 next_node = (current_node - dim_product);
148 // HERE: We use *next* node for calculation (as opposed to current_node!)
149 linkOffset = node_pos_with_loopback_limiter(next_node) + j;
152 XBT_DEBUG("torus_get_route_and_latency - current_node: %lu, next_node: %lu, linkOffset is %lu", current_node,
153 next_node, linkOffset);
157 dim_product *= cur_dim;
160 if (has_limiter()) { // limiter for sender
161 route->link_list_.push_back(get_uplink_from(node_pos_with_loopback(current_node)));
164 resource::StandardLinkImpl* lnk;
166 lnk = get_uplink_from(linkOffset);
168 lnk = get_downlink_to(linkOffset);
170 add_link_latency(route->link_list_, lnk, lat);
172 current_node = next_node;
174 if (has_limiter()) { // limiter for receiver/destination
175 route->link_list_.push_back(get_downlink_to(node_pos_with_loopback(dst->id())));
177 // set gateways (if any)
178 route->gw_src_ = get_gateway(src->id());
179 route->gw_dst_ = get_gateway(dst->id());
182 } // namespace kernel::routing
186 NetZone* create_torus_zone(const std::string& name, const NetZone* parent, const std::vector<unsigned long>& dimensions,
187 const ClusterCallbacks& set_callbacks, double bandwidth, double latency,
188 Link::SharingPolicy sharing_policy)
190 int tot_elements = std::accumulate(dimensions.begin(), dimensions.end(), 1, std::multiplies<>());
191 if (dimensions.empty() || tot_elements <= 0)
192 throw std::invalid_argument("TorusZone: incorrect dimensions parameter, each value must be > 0");
194 throw std::invalid_argument("TorusZone: incorrect bandwidth for internode communication, bw=" +
195 std::to_string(bandwidth));
197 throw std::invalid_argument("TorusZone: incorrect latency for internode communication, lat=" +
198 std::to_string(latency));
200 auto* zone = new kernel::routing::TorusZone(name);
201 zone->set_topology(dimensions);
203 zone->set_parent(parent->get_impl());
205 zone->set_link_characteristics(bandwidth, latency, sharing_policy);
207 for (int i = 0; i < tot_elements; i++) {
208 kernel::routing::NetPoint* netpoint;
211 zone->fill_leaf_from_cb(i, dimensions, set_callbacks, &netpoint, &loopback, &limiter);
213 zone->create_torus_links(netpoint->id(), i, zone->node_pos_with_loopback_limiter(netpoint->id()));
216 return zone->get_iface();
220 } // namespace simgrid