1 /* Copyright (c) 2014-2021. 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/surf/network_interface.hpp"
11 #include <boost/algorithm/string/classification.hpp>
12 #include <boost/algorithm/string/split.hpp>
17 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_route_cluster_torus, surf_route_cluster, "Torus Routing part of surf");
23 void TorusZone::create_torus_links(unsigned long id, int rank, unsigned long position)
25 /* Create all links that exist in the torus. Each rank creates @a dimensions-1 links */
26 unsigned long dim_product = 1; // Needed to calculate the next neighbor_id
28 for (unsigned long j = 0; j < dimensions_.size(); j++) {
29 unsigned long current_dimension =
30 dimensions_[j]; // which dimension are we currently in?
31 // we need to iterate over all dimensions and create all links there
32 // The other node the link connects
33 unsigned long neighbor_rank_id = ((rank / dim_product) % current_dimension == current_dimension - 1)
34 ? rank - (current_dimension - 1) * dim_product
36 // name of neighbor is not right for non contiguous cluster radicals (as id != rank in this case)
37 std::string link_id = get_name() + "_link_from_" + std::to_string(id) + "_to_" + std::to_string(neighbor_rank_id);
38 const s4u::Link* linkup;
39 const s4u::Link* linkdown;
40 if (get_link_sharing_policy() == s4u::Link::SharingPolicy::SPLITDUPLEX) {
41 linkup = create_link(link_id + "_UP", {get_link_bandwidth()})->set_latency(get_link_latency())->seal();
42 linkdown = create_link(link_id + "_DOWN", {get_link_bandwidth()})->set_latency(get_link_latency())->seal();
45 linkup = create_link(link_id, {get_link_bandwidth()})->set_latency(get_link_latency())->seal();
49 * Add the link to its appropriate position.
50 * Note that position rankId*(xbt_dynar_length(dimensions)+has_loopback?+has_limiter?)
51 * holds the link "rankId->rankId"
53 add_private_link_at(position + j, {linkup->get_impl(), linkdown->get_impl()});
54 dim_product *= current_dimension;
58 std::vector<unsigned long> TorusZone::parse_topo_parameters(const std::string& topo_parameters)
60 std::vector<std::string> dimensions_str;
61 boost::split(dimensions_str, topo_parameters, boost::is_any_of(","));
62 std::vector<unsigned long> dimensions;
64 /* We are in a torus cluster
65 * Parse attribute dimensions="dim1,dim2,dim3,...,dimN" and save them into a vector.
66 * Additionally, we need to know how many ranks we have in total
68 std::transform(begin(dimensions_str), end(dimensions_str), std::back_inserter(dimensions),
69 [](const std::string& s) { return std::stoi(s); });
74 void TorusZone::set_topology(const std::vector<unsigned long>& dimensions)
76 xbt_assert(not dimensions.empty(), "Torus dimensions cannot be empty");
77 dimensions_ = dimensions;
78 set_num_links_per_node(dimensions_.size());
81 void TorusZone::get_local_route(const NetPoint* src, const NetPoint* dst, Route* route, double* lat)
83 XBT_VERB("torus getLocalRoute from '%s'[%lu] to '%s'[%lu]", src->get_cname(), src->id(), dst->get_cname(), dst->id());
85 if (dst->is_router() || src->is_router())
88 if (src->id() == dst->id() && has_loopback()) {
89 resource::LinkImpl* uplink = get_uplink_from(node_pos(src->id()));
91 add_link_latency(route->link_list_, uplink, lat);
96 * Dimension based routing routes through each dimension consecutively
97 * TODO Change to dynamic assignment
101 * Arrays that hold the coordinates of the current node and the target; comparing the values at the i-th position of
102 * both arrays, we can easily assess whether we need to route into this dimension or not.
104 const unsigned long dsize = dimensions_.size();
105 std::vector<unsigned long> myCoords(dsize);
106 std::vector<unsigned long> targetCoords(dsize);
107 unsigned int dim_size_product = 1;
108 for (unsigned long i = 0; i < dsize; i++) {
109 unsigned long cur_dim_size = dimensions_[i];
110 myCoords[i] = (src->id() / dim_size_product) % cur_dim_size;
111 targetCoords[i] = (dst->id() / dim_size_product) % cur_dim_size;
112 dim_size_product *= cur_dim_size;
116 * linkOffset describes the offset where the link we want to use is stored(+1 is added because each node has a link
117 * from itself to itself, which can only be the case if src->m_id == dst->m_id -- see above for this special case)
119 unsigned long linkOffset = (dsize + 1) * src->id();
121 bool use_lnk_up = false; // Is this link of the form "cur -> next" or "next -> cur"? false means: next -> cur
122 unsigned long current_node = src->id();
123 while (current_node != dst->id()) {
124 unsigned long next_node = 0;
125 unsigned long dim_product = 1; // First, we will route in x-dimension
126 for (unsigned long j = 0; j < dsize; j++) {
127 const unsigned long cur_dim = dimensions_[j];
128 // current_node/dim_product = position in current dimension
129 if ((current_node / dim_product) % cur_dim != (dst->id() / dim_product) % cur_dim) {
130 if ((targetCoords[j] > myCoords[j] &&
131 targetCoords[j] <= myCoords[j] + cur_dim / 2) // Is the target node on the right, without the wrap-around?
133 (myCoords[j] > cur_dim / 2 && (myCoords[j] + cur_dim / 2) % cur_dim >=
134 targetCoords[j])) { // Or do we need to use the wrap around to reach it?
135 if ((current_node / dim_product) % cur_dim == cur_dim - 1)
136 next_node = (current_node + dim_product - dim_product * cur_dim);
138 next_node = (current_node + dim_product);
140 // HERE: We use *CURRENT* node for calculation (as opposed to next_node)
141 linkOffset = node_pos_with_loopback_limiter(current_node) + j;
143 } else { // Route to the left
144 if ((current_node / dim_product) % cur_dim == 0)
145 next_node = (current_node - dim_product + dim_product * cur_dim);
147 next_node = (current_node - dim_product);
149 // HERE: We use *next* node for calculation (as opposed to current_node!)
150 linkOffset = node_pos_with_loopback_limiter(next_node) + j;
153 XBT_DEBUG("torus_get_route_and_latency - current_node: %lu, next_node: %lu, linkOffset is %lu", current_node,
154 next_node, linkOffset);
158 dim_product *= cur_dim;
161 if (has_limiter()) { // limiter for sender
162 route->link_list_.push_back(get_uplink_from(node_pos_with_loopback(current_node)));
165 resource::LinkImpl* lnk;
167 lnk = get_uplink_from(linkOffset);
169 lnk = get_downlink_to(linkOffset);
171 add_link_latency(route->link_list_, lnk, lat);
173 current_node = next_node;
175 if (has_limiter()) { // limiter for receiver/destination
176 route->link_list_.push_back(get_downlink_to(node_pos_with_loopback(dst->id())));
178 // set gateways (if any)
179 route->gw_src_ = get_gateway(src->id());
180 route->gw_dst_ = get_gateway(dst->id());
183 } // namespace routing
184 } // namespace kernel
188 NetZone* create_torus_zone(const std::string& name, const NetZone* parent, const std::vector<unsigned long>& dimensions,
189 const ClusterCallbacks& set_callbacks, double bandwidth, double latency,
190 Link::SharingPolicy sharing_policy)
192 int tot_elements = std::accumulate(dimensions.begin(), dimensions.end(), 1, std::multiplies<>());
193 if (dimensions.empty() || tot_elements <= 0)
194 throw std::invalid_argument("TorusZone: incorrect dimensions parameter, each value must be > 0");
196 throw std::invalid_argument("TorusZone: incorrect bandwidth for internode communication, bw=" +
197 std::to_string(bandwidth));
199 throw std::invalid_argument("TorusZone: incorrect latency for internode communication, lat=" +
200 std::to_string(latency));
202 auto* zone = new kernel::routing::TorusZone(name);
203 zone->set_topology(dimensions);
205 zone->set_parent(parent->get_impl());
207 zone->set_link_characteristics(bandwidth, latency, sharing_policy);
209 for (int i = 0; i < tot_elements; i++) {
210 kernel::routing::NetPoint* netpoint;
213 zone->fill_leaf_from_cb(i, dimensions, set_callbacks, &netpoint, &loopback, &limiter);
215 zone->create_torus_links(netpoint->id(), i, zone->node_pos_with_loopback_limiter(netpoint->id()));
218 return zone->get_iface();
222 } // namespace simgrid