1 /* Copyright (c) 2009-2018. 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-indepent trace.
18 3. Create inter-process noise, by spawning useless actors.
19 4. Wait for completion (implicitly, via MSG_main's return)
29 #include <boost/algorithm/string.hpp>
31 #include <simgrid/msg.h>
32 #include <simgrid/s4u.hpp>
33 #include <smpi/smpi.h>
34 #include <xbt/file.hpp>
36 XBT_LOG_NEW_DEFAULT_CATEGORY(replay_multiple_manual, "Messages specific for this example");
39 std::string smpi_app_name; //!< The unique name of the SMPI application
40 std::string filename; //!< The filename of the main trace file (which contains other filenames for each rank)
41 int app_size; //!< The number of processes (actors) of the job
42 int starting_time; //!< When the job should start
43 std::vector<int> allocation; //!< Where the job should be executed. Values are hosts indexes.
44 std::vector<std::string> traces_filenames; //!< The filenames of the different action files. Read from filename.
45 int unique_job_number; //!< The job unique number in [0, n[.
48 // ugly globals to avoid creating structures for giving args to processes
49 std::vector<simgrid::s4u::Host*> hosts;
50 int noise_between_jobs;
52 static bool job_comparator(const Job* j1, const Job* j2)
54 if (j1->starting_time == j2->starting_time)
55 return j1->smpi_app_name < j2->smpi_app_name;
56 return j1->starting_time < j2->starting_time;
59 struct s_smpi_replay_process_args {
65 static int smpi_replay_process(int argc, char* argv[])
67 s_smpi_replay_process_args* args = static_cast<s_smpi_replay_process_args*>(MSG_process_get_data(MSG_process_self()));
69 if (args->semaphore != nullptr)
70 MSG_sem_acquire(args->semaphore);
72 XBT_INFO("Replaying rank %d of job %d (smpi_app '%s')", args->rank, args->job->unique_job_number,
73 args->job->smpi_app_name.c_str());
75 smpi_replay_run(&argc, &argv);
76 XBT_INFO("Finished replaying rank %d of job %d (smpi_app '%s')", args->rank, args->job->unique_job_number,
77 args->job->smpi_app_name.c_str());
79 if (args->semaphore != nullptr)
80 MSG_sem_release(args->semaphore);
86 // Sleeps for a given amount of time
87 static int sleeper_process(int* param)
89 XBT_DEBUG("Sleeping for %d seconds", *param);
90 simgrid::s4u::this_actor::sleep_for(*param);
97 // Launches some sleeper processes
98 static void pop_some_processes(int nb_processes, simgrid::s4u::Host* host)
100 for (int i = 0; i < nb_processes; ++i) {
101 int* param = new int;
103 simgrid::s4u::Actor::create("meh", host, sleeper_process, param);
107 static int job_executor_process(Job* job)
109 msg_sem_t job_semaphore = MSG_sem_init(1);
110 XBT_INFO("Executing job %d (smpi_app '%s')", job->unique_job_number, job->smpi_app_name.c_str());
112 for (int i = 0; i < job->app_size; ++i) {
113 char** argv = xbt_new(char*, 5);
114 argv[0] = xbt_strdup("1"); // log only?
115 argv[1] = xbt_strdup(job->smpi_app_name.c_str()); // application instance
116 argv[2] = bprintf("%d", i); // rank
117 argv[3] = xbt_strdup(job->traces_filenames[i].c_str()); // smpi trace file for this rank
118 argv[4] = xbt_strdup("0"); // ?
120 s_smpi_replay_process_args* args = new s_smpi_replay_process_args;
122 args->semaphore = nullptr;
126 args->semaphore = job_semaphore;
128 char* str_pname = bprintf("%d_%d", job->unique_job_number, i);
129 MSG_process_create_with_arguments(str_pname, smpi_replay_process, (void*)args, hosts[job->allocation[i]], 5, argv);
133 MSG_sem_acquire(job_semaphore);
134 MSG_sem_destroy(job_semaphore);
136 XBT_INFO("Finished job %d (smpi_app '%s')", job->unique_job_number, job->smpi_app_name.c_str());
141 // Executes a workload of SMPI processes
142 static int workload_executor_process(std::vector<Job*>* workload)
144 for (Job* job : *workload) {
145 // Let's wait until the job's waiting time if needed
146 double curr_time = simgrid::s4u::Engine::get_clock();
147 if (job->starting_time > curr_time) {
148 double time_to_sleep = (double)job->starting_time - curr_time;
149 XBT_INFO("Sleeping %g seconds (waiting for job %d, app '%s')", time_to_sleep, job->starting_time,
150 job->smpi_app_name.c_str());
151 simgrid::s4u::this_actor::sleep_for(time_to_sleep);
154 if (noise_between_jobs > 0) {
155 // Let's add some process noise
156 XBT_DEBUG("Popping %d noise processes before running job %d (app '%s')", noise_between_jobs,
157 job->unique_job_number, job->smpi_app_name.c_str());
158 pop_some_processes(noise_between_jobs, hosts[job->allocation[0]]);
161 // Let's finally run the job executor
162 std::string job_process_name = "job_" + job->smpi_app_name;
163 XBT_INFO("Launching the job executor of job %d (app '%s')", job->unique_job_number, job->smpi_app_name.c_str());
164 simgrid::s4u::Actor::create(job_process_name.c_str(), hosts[job->allocation[0]], job_executor_process, job);
170 // Reads jobs from a workload file and returns them
171 static std::vector<Job*> all_jobs(const std::string& workload_file)
173 std::ifstream f(workload_file);
174 xbt_assert(f.is_open(), "Cannot open file '%s'.", workload_file.c_str());
175 std::vector<Job*> jobs;
177 simgrid::xbt::Path path(workload_file);
178 std::string dir = path.get_dir_name();
181 while (std::getline(f, line)) {
182 std::string app_name;
183 std::string filename_unprefixed;
188 std::istringstream is(line);
189 if (is >> app_name >> filename_unprefixed >> app_size >> starting_time >> alloc) {
193 job->smpi_app_name = app_name;
194 job->filename = dir + "/" + filename_unprefixed;
195 job->app_size = app_size;
196 job->starting_time = starting_time;
198 std::vector<std::string> subparts;
199 boost::split(subparts, alloc, boost::is_any_of(","), boost::token_compress_on);
201 if ((int)subparts.size() != job->app_size)
202 throw std::runtime_error("size/alloc inconsistency");
204 job->allocation.resize(subparts.size());
205 for (unsigned int i = 0; i < subparts.size(); ++i)
206 job->allocation[i] = stoi(subparts[i]);
208 // Let's read the filename
209 std::ifstream traces_file(job->filename);
210 if (!traces_file.is_open())
211 throw std::runtime_error("Cannot open file " + job->filename);
213 std::string traces_line;
214 while (std::getline(traces_file, traces_line)) {
215 boost::trim_right(traces_line);
216 job->traces_filenames.push_back(dir + "/" + traces_line);
219 if (static_cast<int>(job->traces_filenames.size()) < job->app_size)
220 throw std::runtime_error("size/tracefiles inconsistency");
221 job->traces_filenames.resize(job->app_size);
223 XBT_INFO("Job read: app='%s', file='%s', size=%d, start=%d, "
225 job->smpi_app_name.c_str(), filename_unprefixed.c_str(), job->app_size, job->starting_time,
228 } catch (const std::exception& e) {
229 printf("Bad line '%s' of file '%s': %s.\n", line.c_str(), workload_file.c_str(), e.what());
234 // Jobs are sorted by ascending date, then by lexicographical order of their
236 sort(jobs.begin(), jobs.end(), job_comparator);
238 for (unsigned int i = 0; i < jobs.size(); ++i)
239 jobs[i]->unique_job_number = i;
244 int main(int argc, char* argv[])
247 "Usage: %s platform_file workload_file initial_noise noise_between_jobs\n"
248 "\tExample: %s platform.xml workload_compute\n",
251 // Simulation setting
252 MSG_init(&argc, argv);
253 simgrid::s4u::Engine e(&argc, argv);
254 e.load_platform(argv[1]);
255 hosts = e.get_all_hosts();
256 xbt_assert(hosts.size() >= 4, "The given platform should contain at least 4 hosts (found %zu).", hosts.size());
258 // Let's retrieve all SMPI jobs
259 std::vector<Job*> jobs = all_jobs(argv[2]);
261 // Let's register them
262 for (const Job* job : jobs)
263 SMPI_app_instance_register(job->smpi_app_name.c_str(), smpi_replay_process, job->app_size);
267 // Read noise arguments
268 int initial_noise = std::stoi(argv[3]);
269 xbt_assert(initial_noise >= 0, "Invalid initial_noise argument");
271 noise_between_jobs = std::stoi(argv[4]);
272 xbt_assert(noise_between_jobs >= 0, "Invalid noise_between_jobs argument");
274 if (initial_noise > 0) {
275 XBT_DEBUG("Popping %d noise processes", initial_noise);
276 pop_some_processes(initial_noise, hosts[0]);
279 // Let's execute the workload
280 simgrid::s4u::Actor::create("workload_executor", hosts[0], workload_executor_process, &jobs);
283 XBT_INFO("Simulation finished! Final time: %g", e.get_clock());
287 for (const Job* job : jobs)