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 cluster
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 /* seal newly created netzone */
123 return std::make_pair(host_zone->get_netpoint(), gateway->get_netpoint());
126 /*************************************************************************************************/
128 * @brief Callback to create limiter link (1Gbs) for each netpoint
130 * The coord parameter depends on the cluster being created:
131 * - Torus: Direct translation of the Torus' dimensions, e.g. (0, 0, 0) for a 3-D Torus
132 * - Fat-Tree: A pair (level in the tree, id), e.g. (0, 0) for first leaf in the tree and (1,0) for the first switch at
134 * - Dragonfly: a tuple (group, chassis, blades/routers, nodes), e.g. (0, 0, 0, 0) for first node in the cluster. To
135 * identify the router inside a (group, chassis, blade), we use MAX_UINT in the last parameter (e.g. 0, 0, 0,
138 * @param zone Torus netzone being created (usefull to create the hosts/links inside it)
139 * @param coord Coordinates in the cluster
140 * @param id Internal identifier in the torus (for information)
141 * @return Limiter link
143 static sg4::Link* create_limiter(sg4::NetZone* zone, const std::vector<unsigned int>& /*coord*/, int id)
145 return zone->create_link("limiter-" + std::to_string(id), 1e9)->seal();
149 * @brief Creates a TORUS cluster
151 * Creates a TORUS cluster with dimensions 2x2x2
153 * The cluster has 8 elements/leaves in total. Each element is a StarZone containing 8 Hosts.
154 * Each pair in the torus is connected through 2 links:
155 * 1) limiter: a 1Gbs limiter link (set by user through the set_limiter callback)
156 * 2) link: 10Gbs link connecting the components (created automatically)
165 * B-----+----C (X-axis=2)
167 * For example, a communication from A to C goes through:
168 * <tt> A->limiter(A)->link(A-B)->limiter(B)->link(B-C)->limiter(C)->C </tt>
170 * More precisely, considering that A and C are StarZones, a
171 * communication from A-CPU-3 to C-CPU-7 goes through:
172 * 1) StarZone A: A-CPU-3 -> link-up-A-CPU-3 -> A-CPU-0
173 * 2) A-CPU-0->limiter(A)->link(A-B)->limiter(B)->link(B-C)->limiter(C)->C-CPU-0
174 * 3) StarZone C: C-CPU-0-> link-down-C-CPU-7 -> C-CPU-7
176 * Note that we don't have limiter links inside the StarZones(A, B, C),
177 * but we have limiters in the Torus that are added to the links in the path (as we can see in "2)")
179 * More details in: <a href="https://simgrid.org/doc/latest/Platform_examples.html?highlight=torus#torus-cluster">Torus
182 static void create_torus_cluster()
184 /* create the torus cluster, 10Gbs link between elements in the cluster */
185 sg4::create_torus_zone("cluster", nullptr, {2, 2, 2}, {create_hostzone, {}, create_limiter}, 10e9, 10e-6,
186 sg4::Link::SharingPolicy::SPLITDUPLEX)
190 /*************************************************************************************************/
192 * @brief Creates a Fat-Tree cluster
194 * Creates a Fat-Tree cluster with 2 levels and 6 nodes
195 * The following parameters are used to create this cluster:
196 * - Levels: 2 - two-level of switches in the cluster
197 * - Down links: 2, 3 - L2 routers is connected to 2 elements, L1 routers to 3 elements
198 * - Up links: 1, 2 - Each node (A-F) is connected to 1 L1 router, L1 routers are connected to 2 L2
199 * - Link count: 1, 1 - Use 1 link in each level
201 * The first parameter describes how many levels we have.
202 * The following ones describe the connection between the elements and must have exactly n_levels components.
205 * S3 S4 <-- Level 2 routers
206 * link:limiter - / \ / \
208 * link: 10GBps --> | / \ |
209 * (full-duplex) | / \ |
212 * S1 S2 <-- Level 1 routers
213 * link:limiter -> | |
215 * link:10GBps --> /|\ /|\
218 * link:limiter -> / | \ / | \
219 * A B C D E F <-- level 0 Nodes
221 * Each element (A to F) is a StarZone containing 8 Hosts.
222 * The connection uses 2 links:
223 * 1) limiter: a 1Gbs limiter link (set by user through the set_limiter callback)
224 * 2) link: 10Gbs link connecting the components (created automatically)
226 * For example, a communication from A to C goes through:
227 * <tt> A->limiter(A)->link(A-S1)->limiter(S1)->link(S1-C)->->limiter(C)->C</tt>
229 * More precisely, considering that A and C are StarZones, a
230 * communication from A-CPU-3 to C-CPU-7 goes through:
231 * 1) StarZone A: A-CPU-3 -> link-up-A-CPU-3 -> A-CPU-0
232 * 2) A-CPU-0->limiter(A)->link(A-S1)->limiter(S1)->link(S1-C)->limiter(C)->C-CPU-0
233 * 3) StarZone C: C-CPU-0-> link-down-C-CPU-7 -> C-CPU-7
235 * More details in: <a href="https://simgrid.org/doc/latest/Platform_examples.html#fat-tree-cluster">Fat-Tree
238 static void create_fatTree_cluster()
240 /* create the fat tree cluster, 10Gbs link between elements in the cluster */
241 sg4::create_fatTree_zone("cluster", nullptr, {2, {2, 3}, {1, 2}, {1, 1}}, {create_hostzone, {}, create_limiter}, 10e9,
242 10e-6, sg4::Link::SharingPolicy::SPLITDUPLEX)
246 /*************************************************************************************************/
248 * @brief Creates a Dragonfly cluster
250 * Creates a Dragonfly cluster with 2 groups and 16 nodes
251 * The following parameters are used to create this cluster:
252 * - Groups: 2 groups, connected with 2 links (blue links)
253 * - Chassis: 2 chassis, connected with a single link (black links)
254 * - Routers: 2 routers, connected with 2 links (green links)
255 * - Nodes: 2 leaves per router, single link
257 * The diagram below illustrates a group in the dragonfly cluster
259 * +------------------------------------------------+
261 * | +------------------------+ |
262 * | +---|--------------+ +---|--------------+ |
263 * | | | green | | | green | |
264 * | | | links (2) | | | links (2) | | blue links(2)
265 * | | R1 ====== R2 | | R3 ----- R4 ======================> "Group 2"
266 * | | / \ / \ | | / \ / \ | |
267 * | | A B C D | | E F G H | |
268 * | +------------------+ +------------------+ |
269 * | Chassis 1 Chassis 2 |
270 * +------------------------------------------------+
273 * Each element (A, B, C, etc) is a StarZone containing 8 Hosts.
274 * The connection between elements (e.g. A->R1) uses 2 links:
275 * 1) limiter: a 1Gbs limiter link (set by user through the set_limiter callback)
276 * 2) link: 10Gbs link connecting the components (created automatically)
278 * For example, a communication from A to C goes through:
279 * <tt> A->limiter(A)->link(A-R1)->limiter(R1)->link(R1-R2)->limiter(R2)->link(R2-C)limiter(C)->C</tt>
281 * More details in: <a href="https://simgrid.org/doc/latest/Platform_examples.html#dragonfly-cluster">Dragonfly
284 static void create_dragonfly_cluster()
286 /* create the dragonfly cluster, 10Gbs link between elements in the cluster */
287 sg4::create_dragonfly_zone("cluster", nullptr, {{2, 2}, {2, 1}, {2, 2}, 2}, {create_hostzone, {}, create_limiter},
288 10e9, 10e-6, sg4::Link::SharingPolicy::SPLITDUPLEX)
292 /*************************************************************************************************/
294 int main(int argc, char* argv[])
296 sg4::Engine e(&argc, argv);
297 std::string platform = argv[1];
299 /* create platform */
300 if (platform == "torus")
301 create_torus_cluster();
302 else if (platform == "fatTree")
303 create_fatTree_cluster();
304 else if (platform == "dragonfly")
305 create_dragonfly_cluster();
307 std::vector<sg4::Host*> host_list = e.get_all_hosts();
308 /* create the sender actor running on first host */
309 sg4::Actor::create("sender", host_list[0], Sender(host_list));
310 /* create receiver in every host */
311 for (auto* host : host_list) {
312 sg4::Actor::create(std::string("receiver-") + std::string(host->get_name()), host, Receiver());
315 /* runs the simulation */