1 /* Copyright (c) 2009-2020. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
7 /* This example shows how to replay SMPI time-independent traces in a dynamic
8 fashion. It is inspired from Batsim (https://github.com/oar-team/batsim).
10 The program workflow can be summarized as:
11 1. Read an input workload (set of jobs).
12 Each job is a time-independent trace and a starting time.
13 2. Create initial noise, by spawning useless actors.
14 This is done to avoid SMPI actors to start at actor_id=0.
16 1. Sleep until job's starting time is reached (if needed)
17 2. Launch the replay of the corresponding time-independent trace.
18 3. Create inter-process noise, by spawning useless actors.
19 4. Wait for completion (via s4u::Engine's run method)
29 #include <boost/algorithm/string.hpp>
31 #include <simgrid/s4u.hpp>
32 #include <smpi/smpi.h>
33 #include <xbt/file.hpp>
35 XBT_LOG_NEW_DEFAULT_CATEGORY(replay_multiple_manual, "Messages specific for this example");
38 std::string smpi_app_name; //!< The unique name of the SMPI application
39 std::string filename; //!< The filename of the main trace file (which contains other filenames for each rank)
40 int app_size; //!< The number of processes (actors) of the job
41 int starting_time; //!< When the job should start
42 std::vector<int> allocation; //!< Where the job should be executed. Values are hosts indexes.
43 std::vector<std::string> traces_filenames; //!< The filenames of the different action files. Read from filename.
44 int unique_job_number; //!< The job unique number in [0, n[.
47 // ugly globals to avoid creating structures for giving args to processes
48 static std::vector<simgrid::s4u::Host*> hosts;
49 static int noise_between_jobs;
51 static bool job_comparator(const std::unique_ptr<Job>& j1, const std::unique_ptr<Job>& j2)
53 if (j1->starting_time == j2->starting_time)
54 return j1->smpi_app_name < j2->smpi_app_name;
55 return j1->starting_time < j2->starting_time;
58 static void smpi_replay_process(Job* job, simgrid::s4u::BarrierPtr barrier, int rank)
60 XBT_INFO("Replaying rank %d of job %d (smpi_app '%s')", rank, job->unique_job_number, job->smpi_app_name.c_str());
61 smpi_replay_run(job->smpi_app_name.c_str(), rank, 0, job->traces_filenames[rank].c_str());
62 XBT_INFO("Finished replaying rank %d of job %d (smpi_app '%s')", rank, job->unique_job_number,
63 job->smpi_app_name.c_str());
68 // Sleeps for a given amount of time
69 static int sleeper_process(int param)
71 XBT_DEBUG("Sleeping for %d seconds", param);
72 simgrid::s4u::this_actor::sleep_for(param);
76 // Launches some sleeper processes
77 static void pop_some_processes(int nb_processes, simgrid::s4u::Host* host)
79 for (int i = 0; i < nb_processes; ++i) {
81 simgrid::s4u::Actor::create("meh", host, sleeper_process, param);
85 static int job_executor_process(Job* job)
87 XBT_INFO("Executing job %d (smpi_app '%s')", job->unique_job_number, job->smpi_app_name.c_str());
89 simgrid::s4u::BarrierPtr barrier = simgrid::s4u::Barrier::create(job->app_size + 1);
91 for (int i = 0; i < job->app_size; ++i) {
92 char* str_pname = bprintf("rank_%d_%d", job->unique_job_number, i);
93 simgrid::s4u::Actor::create(str_pname, hosts[job->allocation[i]], smpi_replay_process, job, barrier, i);
99 simgrid::s4u::this_actor::sleep_for(1);
100 XBT_INFO("Finished job %d (smpi_app '%s')", job->unique_job_number, job->smpi_app_name.c_str());
105 // Executes a workload of SMPI processes
106 static int workload_executor_process(const std::vector<std::unique_ptr<Job>>* workload)
108 for (auto const& job : *workload) {
109 // Let's wait until the job's waiting time if needed
110 double curr_time = simgrid::s4u::Engine::get_clock();
111 if (job->starting_time > curr_time) {
112 double time_to_sleep = (double)job->starting_time - curr_time;
113 XBT_INFO("Sleeping %g seconds (waiting for job %d, app '%s')", time_to_sleep, job->starting_time,
114 job->smpi_app_name.c_str());
115 simgrid::s4u::this_actor::sleep_for(time_to_sleep);
118 if (noise_between_jobs > 0) {
119 // Let's add some process noise
120 XBT_DEBUG("Popping %d noise processes before running job %d (app '%s')", noise_between_jobs,
121 job->unique_job_number, job->smpi_app_name.c_str());
122 pop_some_processes(noise_between_jobs, hosts[job->allocation[0]]);
125 // Let's finally run the job executor
126 char* str_pname = bprintf("job_%04d", job->unique_job_number);
127 XBT_INFO("Launching the job executor of job %d (app '%s')", job->unique_job_number, job->smpi_app_name.c_str());
128 simgrid::s4u::Actor::create(str_pname, hosts[job->allocation[0]], job_executor_process, job.get());
135 // Reads jobs from a workload file and returns them
136 static std::vector<std::unique_ptr<Job>> all_jobs(const std::string& workload_file)
138 std::ifstream f(workload_file);
139 xbt_assert(f.is_open(), "Cannot open file '%s'.", workload_file.c_str());
140 std::vector<std::unique_ptr<Job>> jobs;
142 simgrid::xbt::Path path(workload_file);
143 std::string dir = path.get_dir_name();
146 while (std::getline(f, line)) {
147 std::string app_name;
148 std::string filename_unprefixed;
153 std::istringstream is(line);
154 if (is >> app_name >> filename_unprefixed >> app_size >> starting_time >> alloc) {
156 auto job = std::make_unique<Job>();
157 job->smpi_app_name = app_name;
158 job->filename = dir + "/" + filename_unprefixed;
159 job->app_size = app_size;
160 job->starting_time = starting_time;
162 std::vector<std::string> subparts;
163 boost::split(subparts, alloc, boost::is_any_of(","), boost::token_compress_on);
165 if ((int)subparts.size() != job->app_size)
166 throw std::invalid_argument("size/alloc inconsistency");
168 job->allocation.resize(subparts.size());
169 for (unsigned int i = 0; i < subparts.size(); ++i)
170 job->allocation[i] = stoi(subparts[i]);
172 // Let's read the filename
173 std::ifstream traces_file(job->filename);
174 if (!traces_file.is_open())
175 throw std::invalid_argument("Cannot open file " + job->filename);
177 std::string traces_line;
178 while (std::getline(traces_file, traces_line)) {
179 boost::trim_right(traces_line);
180 job->traces_filenames.push_back(dir + "/" + traces_line);
183 if (static_cast<int>(job->traces_filenames.size()) < job->app_size)
184 throw std::invalid_argument("size/tracefiles inconsistency");
185 job->traces_filenames.resize(job->app_size);
187 XBT_INFO("Job read: app='%s', file='%s', size=%d, start=%d, "
189 job->smpi_app_name.c_str(), filename_unprefixed.c_str(), job->app_size, job->starting_time,
191 jobs.emplace_back(std::move(job));
192 } catch (const std::invalid_argument& e) {
193 xbt_die("Bad line '%s' of file '%s': %s.\n", line.c_str(), workload_file.c_str(), e.what());
198 // Jobs are sorted by ascending date, then by lexicographical order of their
200 sort(jobs.begin(), jobs.end(), job_comparator);
202 for (unsigned int i = 0; i < jobs.size(); ++i)
203 jobs[i]->unique_job_number = i;
208 int main(int argc, char* argv[])
211 "Usage: %s platform_file workload_file initial_noise noise_between_jobs\n"
212 "\tExample: %s platform.xml workload_compute\n",
215 // Simulation setting
216 simgrid::s4u::Engine e(&argc, argv);
217 e.load_platform(argv[1]);
218 hosts = e.get_all_hosts();
219 xbt_assert(hosts.size() >= 4, "The given platform should contain at least 4 hosts (found %zu).", hosts.size());
221 // Let's retrieve all SMPI jobs
222 std::vector<std::unique_ptr<Job>> jobs = all_jobs(argv[2]);
224 // Let's register them
225 for (auto const& job : jobs)
226 SMPI_app_instance_register(job->smpi_app_name.c_str(), nullptr, job->app_size);
230 // Read noise arguments
231 int initial_noise = std::stoi(argv[3]);
232 xbt_assert(initial_noise >= 0, "Invalid initial_noise argument");
234 noise_between_jobs = std::stoi(argv[4]);
235 xbt_assert(noise_between_jobs >= 0, "Invalid noise_between_jobs argument");
237 if (initial_noise > 0) {
238 XBT_DEBUG("Popping %d noise processes", initial_noise);
239 pop_some_processes(initial_noise, hosts[0]);
242 // Let's execute the workload
243 simgrid::s4u::Actor::create("workload", hosts[0], workload_executor_process, &jobs);
246 XBT_INFO("Simulation finished! Final time: %g", e.get_clock());