1 /* Copyright (c) 2009-2022. 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 /* simple test to schedule a DAX file with the Min-Min algorithm. */
8 #include <simgrid/host.h>
9 #include <simgrid/s4u.hpp>
12 XBT_LOG_NEW_DEFAULT_CATEGORY(dag_scheduling, "Logging specific to this example");
13 namespace sg4 = simgrid::s4u;
15 struct HostAttribute {
16 /* Earliest time at which a host is ready to execute a task */
17 double available_at = 0.0;
18 sg4::Exec* last_scheduled_task = nullptr;
21 static double sg_host_get_available_at(const sg4::Host* host)
23 return host->get_data<HostAttribute>()->available_at;
26 static void sg_host_set_available_at(const sg4::Host* host, double time)
28 host->get_data<HostAttribute>()->available_at = time;
31 static sg4::Exec* sg_host_get_last_scheduled_task(const sg4::Host* host)
33 return host->get_data<HostAttribute>()->last_scheduled_task;
36 static void sg_host_set_last_scheduled_task(const sg4::Host* host, sg4::ExecPtr task)
38 host->get_data<HostAttribute>()->last_scheduled_task = task.get();
41 static bool dependency_exists(const sg4::Exec* src, sg4::Exec* dst)
43 const auto& dependencies = src->get_dependencies();
44 const auto& successors = src->get_successors();
45 return (std::find(successors.begin(), successors.end(), dst) != successors.end() ||
46 dependencies.find(dst) != dependencies.end());
49 static std::vector<sg4::Exec*> get_ready_tasks(const std::vector<sg4::ActivityPtr>& dax)
51 std::vector<sg4::Exec*> ready_tasks;
52 std::map<sg4::Exec*, unsigned int> candidate_execs;
55 // Only look at activity that have their dependencies solved but are not assigned
56 if (a->dependencies_solved() && not a->is_assigned()) {
57 // if it is an exec, it's ready
58 if (auto* exec = dynamic_cast<sg4::Exec*>(a.get()))
59 ready_tasks.push_back(exec);
60 // if it a comm, we consider its successor as a candidate. If a candidate solves all its dependencies,
61 // i.e., get all its input data, it's ready
62 if (const auto* comm = dynamic_cast<sg4::Comm*>(a.get())) {
63 auto* next_exec = static_cast<sg4::Exec*>(comm->get_successors().front().get());
64 candidate_execs[next_exec]++;
65 if (next_exec->get_dependencies().size() == candidate_execs[next_exec])
66 ready_tasks.push_back(next_exec);
70 XBT_DEBUG("There are %zu ready tasks", ready_tasks.size());
74 static double finish_on_at(const sg4::ExecPtr task, const sg4::Host* host)
78 const auto& parents = task->get_dependencies();
80 if (not parents.empty()) {
81 double data_available = 0.;
82 double last_data_available;
83 /* compute last_data_available */
84 last_data_available = -1.0;
85 for (const auto& parent : parents) {
87 if (const auto* comm = dynamic_cast<sg4::Comm*>(parent.get())) {
88 auto source = comm->get_source();
89 XBT_DEBUG("transfer from %s to %s", source->get_cname(), host->get_cname());
90 /* Estimate the redistribution time from this parent */
92 if (comm->get_remaining() <= 1e-6) {
95 redist_time = sg_host_get_route_latency(source, host) +
96 comm->get_remaining() / sg_host_get_route_bandwidth(source, host);
98 // We use the user data field to store the finish time of the predecessor of the comm, i.e., its potential start
100 data_available = *comm->get_data<double>() + redist_time;
103 /* no transfer, control dependency */
104 if (const auto* exec = dynamic_cast<sg4::Exec*>(parent.get())) {
105 data_available = exec->get_finish_time();
108 if (last_data_available < data_available)
109 last_data_available = data_available;
112 result = fmax(sg_host_get_available_at(host), last_data_available) + task->get_remaining() / host->get_speed();
114 result = sg_host_get_available_at(host) + task->get_remaining() / host->get_speed();
119 static sg4::Host* get_best_host(const sg4::ExecPtr exec)
121 std::vector<sg4::Host*> hosts = sg4::Engine::get_instance()->get_all_hosts();
122 auto best_host = hosts.front();
123 double min_EFT = finish_on_at(exec, best_host);
125 for (const auto& h : hosts) {
126 double EFT = finish_on_at(exec, h);
127 XBT_DEBUG("%s finishes on %s at %f", exec->get_cname(), h->get_cname(), EFT);
137 static void schedule_on(sg4::ExecPtr exec, sg4::Host* host)
139 exec->set_host(host);
140 // we can also set the destination of all the input comms of this exec
141 for (const auto& pred : exec->get_dependencies()) {
142 auto* comm = dynamic_cast<sg4::Comm*>(pred.get());
143 if (comm != nullptr) {
144 comm->set_destination(host);
145 delete comm->get_data<double>();
148 // we can also set the source of all the output comms of this exec
149 for (const auto& succ : exec->get_successors()) {
150 auto* comm = dynamic_cast<sg4::Comm*>(succ.get());
152 comm->set_source(host);
156 int main(int argc, char** argv)
158 sg4::Engine e(&argc, argv);
159 std::set<sg4::Activity*> vetoed;
160 e.track_vetoed_activities(&vetoed);
162 sg4::Activity::on_completion_cb([](sg4::Activity const& activity) {
163 // when an Exec completes, we need to set the potential start time of all its ouput comms
164 const auto* exec = dynamic_cast<sg4::Exec const*>(&activity);
165 if (exec == nullptr) // Only Execs are concerned here
167 for (const auto& succ : exec->get_successors()) {
168 auto* comm = dynamic_cast<sg4::Comm*>(succ.get());
169 if (comm != nullptr) {
170 auto* finish_time = new double(exec->get_finish_time());
171 // We use the user data field to store the finish time of the predecessor of the comm, i.e., its potential start
173 comm->set_data(finish_time);
178 e.load_platform(argv[1]);
180 /* Allocating the host attribute */
181 unsigned long total_nhosts = e.get_host_count();
182 const auto hosts = e.get_all_hosts();
183 std::vector<HostAttribute> host_attributes(total_nhosts);
184 for (unsigned long i = 0; i < total_nhosts; i++)
185 hosts[i]->set_data(&host_attributes[i]);
187 /* load the DAX file */
188 auto dax = sg4::create_DAG_from_DAX(argv[2]);
190 /* Schedule the root first */
191 auto* root = static_cast<sg4::Exec*>(dax.front().get());
192 auto host = get_best_host(root);
193 schedule_on(root, host);
197 while (not vetoed.empty()) {
198 XBT_DEBUG("Start new scheduling round");
199 /* Get the set of ready tasks */
200 auto ready_tasks = get_ready_tasks(dax);
203 if (ready_tasks.empty()) {
204 /* there is no ready task, let advance the simulation */
208 /* For each ready task:
209 * get the host that minimizes the completion time.
210 * select the task that has the minimum completion time on its best host.
212 double min_finish_time = -1.0;
213 sg4::Exec* selected_task = nullptr;
214 sg4::Host* selected_host = nullptr;
216 for (auto task : ready_tasks) {
217 XBT_DEBUG("%s is ready", task->get_cname());
218 host = get_best_host(task);
219 double finish_time = finish_on_at(task, host);
220 if (min_finish_time < 0 || finish_time < min_finish_time) {
221 min_finish_time = finish_time;
222 selected_task = task;
223 selected_host = host;
227 XBT_INFO("Schedule %s on %s", selected_task->get_cname(), selected_host->get_cname());
228 schedule_on(selected_task, selected_host);
231 * tasks can be executed concurrently when they can by default.
232 * Yet schedulers take decisions assuming that tasks wait for resource availability to start.
233 * The solution (well crude hack is to keep track of the last task scheduled on a host and add a special type of
234 * dependency if needed to force the sequential execution meant by the scheduler.
235 * If the last scheduled task is already done, has failed or is a predecessor of the current task, no need for a
239 auto last_scheduled_task = sg_host_get_last_scheduled_task(selected_host);
240 if (last_scheduled_task && (last_scheduled_task->get_state() != sg4::Activity::State::FINISHED) &&
241 (last_scheduled_task->get_state() != sg4::Activity::State::FAILED) &&
242 not dependency_exists(sg_host_get_last_scheduled_task(selected_host), selected_task))
243 last_scheduled_task->add_successor(selected_task);
245 sg_host_set_last_scheduled_task(selected_host, selected_task);
246 sg_host_set_available_at(selected_host, min_finish_time);
252 XBT_INFO("Simulation Time: %f", simgrid_get_clock());