1 /* Copyright (c) 2010-2021. 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 /* This example shows how to build a torus cluster with multi-core hosts.
8 * However, each leaf in the torus is a StarZone, composed of several CPUs
10 * Each actor runs in a specific CPU. One sender broadcasts a message to all receivers.
13 #include "simgrid/s4u.hpp"
14 namespace sg4 = simgrid::s4u;
16 XBT_LOG_NEW_DEFAULT_CATEGORY(s4u_torus_multicpu, "Messages specific for this s4u example");
19 long msg_size = 1e6; /* message size in bytes */
20 std::vector<sg4::Host*> hosts_;
23 explicit Sender(const std::vector<sg4::Host*> hosts) : hosts_{hosts} {}
24 void operator()() const
26 /* Vector in which we store all ongoing communications */
27 std::vector<sg4::CommPtr> pending_comms;
29 /* Make a vector of the mailboxes to use */
30 std::vector<sg4::Mailbox*> mboxes;
32 /* Start dispatching 1 message to all receivers */
33 std::string msg_content =
34 std::string("Hello, I'm alive and running on ") + std::string(sg4::this_actor::get_host()->get_name());
35 for (const auto* host : hosts_) {
36 /* Copy the data we send: the 'msg_content' variable is not a stable storage location.
37 * It will be destroyed when this actor leaves the loop, ie before the receiver gets it */
38 auto* payload = new std::string(msg_content);
40 /* Create a communication representing the ongoing communication, and store it in pending_comms */
41 auto mbox = sg4::Mailbox::by_name(host->get_name());
42 mboxes.push_back(mbox);
43 sg4::CommPtr comm = mbox->put_async(payload, msg_size);
44 pending_comms.push_back(comm);
47 XBT_INFO("Done dispatching all messages");
49 /* Now that all message exchanges were initiated, wait for their completion in one single call */
50 sg4::Comm::wait_all(&pending_comms);
52 XBT_INFO("Goodbye now!");
56 /* Receiver actor: wait for 1 message on the mailbox identified by the hostname */
59 void operator()() const
61 auto mbox = sg4::Mailbox::by_name(sg4::this_actor::get_host()->get_name());
62 auto received = mbox->get_unique<std::string>();
63 XBT_INFO("I got a '%s'.", received->c_str());
67 /*************************************************************************************************/
69 * @brief Callback to set a cluster leaf/element
71 * In our example, each leaf if a StarZone, composed of 8 CPUs.
72 * Each CPU is modeled as a host, connected to the outer world through a high-speed PCI link.
73 * Obs.: CPU0 is the gateway for this zone
81 * / / \ \<-- 100Gbs, 10us link (1 link UP and 1 link DOWN for full-duplex)
86 * @param zone Cluster netzone being created (usefull to create the hosts/links inside it)
87 * @param coord Coordinates in the torus (e.g. "0,0,0", "0,1,0")
88 * @param id Internal identifier in the torus (for information)
89 * @return netpoint, gateway: the netpoint to the StarZone and CPU0 as gateway
91 static std::pair<simgrid::kernel::routing::NetPoint*, simgrid::kernel::routing::NetPoint*>
92 create_hostzone(const sg4::NetZone* zone, const std::vector<unsigned int>& /*coord*/, int id)
94 constexpr int num_cpus = 8; //!< Number of CPUs in the zone
95 constexpr double speed = 1e9; //!< Speed of each CPU
96 constexpr double link_bw = 100e9; //!< Link bw connecting the CPU
97 constexpr double link_lat = 1e-9; //!< Link latency
99 std::string hostname = "host" + std::to_string(id);
100 /* create the StarZone */
101 auto* host_zone = sg4::create_star_zone(hostname);
102 /* setting my Torus parent zone */
103 host_zone->set_parent(zone);
105 const sg4::Host* gateway = nullptr;
107 for (int i = 0; i < num_cpus; i++) {
108 std::string cpu_name = hostname + "-cpu" + std::to_string(i);
109 const sg4::Host* host = host_zone->create_host(cpu_name, speed)->seal();
110 /* the first CPU is the gateway */
113 /* create 2 links for a full-duplex communication */
114 sg4::Link* link_up = host_zone->create_link("link-up-" + cpu_name, link_bw)->set_latency(link_lat)->seal();
115 sg4::Link* link_down = host_zone->create_link("link-down-" + cpu_name, link_bw)->set_latency(link_lat)->seal();
116 /* link UP, connection from CPU to outer world */
117 host_zone->add_route(host->get_netpoint(), nullptr, nullptr, nullptr, std::vector<sg4::Link*>{link_up}, false);
118 /* link DOWN, connection from outer to CPU */
119 host_zone->add_route(nullptr, host->get_netpoint(), nullptr, nullptr, std::vector<sg4::Link*>{link_down}, false);
121 return std::make_pair(host_zone->get_netpoint(), gateway->get_netpoint());
124 /*************************************************************************************************/
126 * @brief Callback to create limiter link (1Gbs) for each netpoint
127 * @param zone Torus netzone being created (usefull to create the hosts/links inside it)
128 * @param coord Coordinates in the torus (e.g. "0,0,0", "0,1,0")
129 * @param id Internal identifier in the torus (for information)
130 * @return Limiter link
132 static sg4::Link* create_limiter(sg4::NetZone* zone, const std::vector<unsigned int>& coord, int id)
134 return zone->create_link("limiter-" + std::to_string(id), 1e9)->seal();
138 * @brief Creates a TORUS cluster
140 * Creates a TORUS cluster with dimensions 2x2x2
142 * The cluster has 8 elements/leaves in total. Each element is a StarZone containing 8 Hosts.
143 * Each pair in the torus is connected through 2 links:
144 * 1) limiter: a 1Gbs limiter link (set by user through the set_limiter callback)
145 * 2) link: 10Gbs link connecting the components (created automatically)
154 * B-----+----C (X-axis=2)
156 * For example, a communication from A to C goes through:
157 * <tt> A->limiter(A)->link(A-B)->limiter(B)->link(B-C)->limiter(C)->C </tt>
159 * More precisely, considering that A and C are StarZones, a
160 * communication from A-CPU-3 to C-CPU-7 goes through:
161 * 1) StarZone A: A-CPU-3 -> link-up-A-CPU-3 -> A-CPU-0
162 * 2) A-CPU-0->limiter(A)->link(A-B)->limiter(B)->link(B-C)->limiter(C)->C-CPU-0
163 * 3) StarZone C: C-CPU-0-> link-down-C-CPU-7 -> C-CPU-7
165 * Note that we don't have limiter links inside the StarZones(A, B, C),
166 * but we have limiters in the Torus that are added to the links in the path (as we can see in "2)")
168 * More details in: <a href="https://simgrid.org/doc/latest/Platform_examples.html?highlight=torus#torus-cluster">Torus
171 static void create_torus_cluster()
173 /* create the torus cluster, 10Gbs link between elements in the cluster */
174 sg4::create_torus_zone("cluster", nullptr, {2, 2, 2}, {create_hostzone, {}, create_limiter}, 10e9, 10e-6,
175 sg4::Link::SharingPolicy::SPLITDUPLEX)
179 /*************************************************************************************************/
181 * @brief Creates a Fat-Tree cluster
183 * Creates a Fat-Tree cluster with 2 levels and 6 nodes
184 * The following parameters are used to create this cluster:
185 * - Levels: 2 - two-level cluster
186 * - Down links: 2, 3 - L1 routers is connected to 2 elements, L2 routers to 3 elements
187 * - Up links: 1, 2 - Each node (A-F) is connected to 1 L2 router, L2 routers are connected to 2 L1
188 * - Link count: 1, 1 - Use 1 link in each level
190 * The first parameter describes how many levels we have.
191 * The following ones describe the connection between the elements and must have exactly n_levels components.
194 * S3 S4 <-- Level 1 routers
195 * link:limiter - / \ / \
197 * link: 10GBps --> | / \ |
198 * (full-duplex) | / \ |
201 * S1 S2 <-- Level 2 routers
202 * link:limiter -> | |
204 * link:10GBps --> /|\ /|\
207 * link:limiter -> / | \ / | \
208 * A B C D E F <-- Nodes
210 * Each element (A to F) is a StarZone containing 8 Hosts.
211 * The connection uses 2 links:
212 * 1) limiter: a 1Gbs limiter link (set by user through the set_limiter callback)
213 * 2) link: 10Gbs link connecting the components (created automatically)
215 * For example, a communication from A to C goes through:
216 * <tt> A->limiter(A)->link(A-S1)->limiter(S1)->link(S1-C)->->limiter(C)->C</tt>
218 * More precisely, considering that A and C are StarZones, a
219 * communication from A-CPU-3 to C-CPU-7 goes through:
220 * 1) StarZone A: A-CPU-3 -> link-up-A-CPU-3 -> A-CPU-0
221 * 2) A-CPU-0->limiter(A)->link(A-S1)->limiter(S1)->link(S1-C)->limiter(C)->C-CPU-0
222 * 3) StarZone C: C-CPU-0-> link-down-C-CPU-7 -> C-CPU-7
224 * More details in: <a href="https://simgrid.org/doc/latest/Platform_examples.html#fat-tree-cluster">Fat-Tree
227 static void create_fatTree_cluster()
229 /* create the fat tree cluster, 10Gbs link between elements in the cluster */
230 sg4::create_fatTree_zone("cluster", nullptr, {2, {2, 3}, {1, 2}, {1, 1}}, {create_hostzone, {}, create_limiter}, 10e9,
231 10e-6, sg4::Link::SharingPolicy::SPLITDUPLEX)
235 /*************************************************************************************************/
237 * @brief Creates a Dragonfly cluster
239 * Creates a Dragonfly cluster with 2 groups and 16 nodes
240 * The following parameters are used to create this cluster:
241 * - Groups: 2 groups, connected with 2 links (blue links)
242 * - Chassis: 2 chassis, connected with a single link (black links)
243 * - Routers: 2 routers, connected with 2 links (green links)
244 * - Nodes: 2 leaves per router, single link
246 * The diagram below illustrates a group in the dragonfly cluster
248 * +------------------------------------------------+
250 * | +------------------------+ |
251 * | +---|--------------+ +---|--------------+ |
252 * | | | green | | | green | |
253 * | | | links (2) | | | links (2) | | blue links(2)
254 * | | R1 ====== R2 | | R3 ----- R4 ======================> "Group 2"
255 * | | / \ / \ | | / \ / \ | |
256 * | | A B C D | | E F G H | |
257 * | +------------------+ +------------------+ |
258 * | Chassis 1 Chassis 2 |
259 * +------------------------------------------------+
262 * Each element (A, B, C, etc) is a StarZone containing 8 Hosts.
263 * The connection between nodes and routers (e.g. A->R1) uses 2 links:
264 * 1) limiter: a 1Gbs limiter link (set by user through the set_limiter callback)
265 * 2) link: 10Gbs link connecting the components (created automatically)
267 * For example, a communication from A to C goes through:
268 * <tt> A->limiter(A)->link(A-R1)->link(R1-R2)->limiter(C)->C</tt>
270 * Note that limiters are only valid for leaves, not routers.
272 * More details in: <a href="https://simgrid.org/doc/latest/Platform_examples.html#dragonfly-cluster">Dragonfly
275 static void create_dragonfly_cluster()
277 /* create the dragonfly cluster, 10Gbs link between elements in the cluster */
278 sg4::create_dragonfly_zone("cluster", nullptr, {{2, 2}, {2, 1}, {2, 2}, 2}, {create_hostzone, {}, create_limiter},
279 10e9, 10e-6, sg4::Link::SharingPolicy::SPLITDUPLEX)
283 /*************************************************************************************************/
285 int main(int argc, char* argv[])
287 sg4::Engine e(&argc, argv);
288 std::string platform = argv[1];
290 /* create platform */
291 if (platform == "torus")
292 create_torus_cluster();
293 else if (platform == "fatTree")
294 create_fatTree_cluster();
295 else if (platform == "dragonfly")
296 create_dragonfly_cluster();
298 std::vector<sg4::Host*> host_list = e.get_all_hosts();
299 /* create the sender actor running on first host */
300 sg4::Actor::create("sender", host_list[0], Sender(host_list));
301 /* create receiver in every host */
302 for (auto* host : host_list) {
303 sg4::Actor::create(std::string("receiver-") + std::string(host->get_name()), host, Receiver());
306 /* runs the simulation */