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 (via s4u::Engine's run method)
28 #include <boost/algorithm/string.hpp>
30 #include <simgrid/s4u.hpp>
31 #include <smpi/smpi.h>
32 #include <xbt/file.hpp>
34 XBT_LOG_NEW_DEFAULT_CATEGORY(replay_multiple_manual, "Messages specific for this example");
37 std::string smpi_app_name; //!< The unique name of the SMPI application
38 std::string filename; //!< The filename of the main trace file (which contains other filenames for each rank)
39 int app_size; //!< The number of processes (actors) of the job
40 int starting_time; //!< When the job should start
41 std::vector<int> allocation; //!< Where the job should be executed. Values are hosts indexes.
42 std::vector<std::string> traces_filenames; //!< The filenames of the different action files. Read from filename.
43 int unique_job_number; //!< The job unique number in [0, n[.
46 // ugly globals to avoid creating structures for giving args to processes
47 static std::vector<simgrid::s4u::Host*> hosts;
48 static int noise_between_jobs;
50 static bool job_comparator(const Job* j1, const Job* j2)
52 if (j1->starting_time == j2->starting_time)
53 return j1->smpi_app_name < j2->smpi_app_name;
54 return j1->starting_time < j2->starting_time;
57 static void smpi_replay_process(Job* job, simgrid::s4u::BarrierPtr barrier, int rank)
59 // Prepare data for smpi_replay_run
61 char** argv = xbt_new(char*, argc);
62 argv[0] = xbt_strdup("1"); // log only?
63 argv[1] = xbt_strdup(job->smpi_app_name.c_str()); // application instance
64 argv[2] = bprintf("%d", rank); // rank
65 argv[3] = xbt_strdup(job->traces_filenames[rank].c_str()); // smpi trace file for this rank
66 argv[4] = xbt_strdup("0"); // ?
68 // Ugly double storage used for memory deallocation, as SMPI changes argc/argv without cleaning memory.
69 char* to_free[2] = {argv[0], argv[1]}; // <-- This ugly array should disappear.
71 XBT_INFO("Replaying rank %d of job %d (smpi_app '%s')", rank, job->unique_job_number, job->smpi_app_name.c_str());
72 smpi_replay_run(&argc, &argv);
73 XBT_INFO("Finished replaying rank %d of job %d (smpi_app '%s')", rank, job->unique_job_number,
74 job->smpi_app_name.c_str());
79 for (int i = 0; i < 2; ++i) // <-- This ugly loop should disappear.
81 for (int i = 0; i < argc; ++i)
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 XBT_INFO("Executing job %d (smpi_app '%s')", job->unique_job_number, job->smpi_app_name.c_str());
111 simgrid::s4u::BarrierPtr barrier = simgrid::s4u::Barrier::create(job->app_size + 1);
113 for (int i = 0; i < job->app_size; ++i) {
114 char* str_pname = bprintf("%d_%d", job->unique_job_number, i);
115 simgrid::s4u::Actor::create(str_pname, hosts[job->allocation[i]], smpi_replay_process, job, barrier, i);
121 XBT_INFO("Finished job %d (smpi_app '%s')", job->unique_job_number, job->smpi_app_name.c_str());
126 // Executes a workload of SMPI processes
127 static int workload_executor_process(std::vector<Job*>* workload)
129 for (Job* job : *workload) {
130 // Let's wait until the job's waiting time if needed
131 double curr_time = simgrid::s4u::Engine::get_clock();
132 if (job->starting_time > curr_time) {
133 double time_to_sleep = (double)job->starting_time - curr_time;
134 XBT_INFO("Sleeping %g seconds (waiting for job %d, app '%s')", time_to_sleep, job->starting_time,
135 job->smpi_app_name.c_str());
136 simgrid::s4u::this_actor::sleep_for(time_to_sleep);
139 if (noise_between_jobs > 0) {
140 // Let's add some process noise
141 XBT_DEBUG("Popping %d noise processes before running job %d (app '%s')", noise_between_jobs,
142 job->unique_job_number, job->smpi_app_name.c_str());
143 pop_some_processes(noise_between_jobs, hosts[job->allocation[0]]);
146 // Let's finally run the job executor
147 std::string job_process_name = "job_" + job->smpi_app_name;
148 XBT_INFO("Launching the job executor of job %d (app '%s')", job->unique_job_number, job->smpi_app_name.c_str());
149 simgrid::s4u::Actor::create(job_process_name.c_str(), hosts[job->allocation[0]], job_executor_process, job);
155 // Reads jobs from a workload file and returns them
156 static std::vector<Job*> all_jobs(const std::string& workload_file)
158 std::ifstream f(workload_file);
159 xbt_assert(f.is_open(), "Cannot open file '%s'.", workload_file.c_str());
160 std::vector<Job*> jobs;
162 simgrid::xbt::Path path(workload_file);
163 std::string dir = path.get_dir_name();
166 while (std::getline(f, line)) {
167 std::string app_name;
168 std::string filename_unprefixed;
173 std::istringstream is(line);
174 if (is >> app_name >> filename_unprefixed >> app_size >> starting_time >> alloc) {
177 job.smpi_app_name = app_name;
178 job.filename = dir + "/" + filename_unprefixed;
179 job.app_size = app_size;
180 job.starting_time = starting_time;
182 std::vector<std::string> subparts;
183 boost::split(subparts, alloc, boost::is_any_of(","), boost::token_compress_on);
185 if ((int)subparts.size() != job.app_size)
186 throw std::invalid_argument("size/alloc inconsistency");
188 job.allocation.resize(subparts.size());
189 for (unsigned int i = 0; i < subparts.size(); ++i)
190 job.allocation[i] = stoi(subparts[i]);
192 // Let's read the filename
193 std::ifstream traces_file(job.filename);
194 if (!traces_file.is_open())
195 throw std::invalid_argument("Cannot open file " + job.filename);
197 std::string traces_line;
198 while (std::getline(traces_file, traces_line)) {
199 boost::trim_right(traces_line);
200 job.traces_filenames.push_back(dir + "/" + traces_line);
203 if (static_cast<int>(job.traces_filenames.size()) < job.app_size)
204 throw std::invalid_argument("size/tracefiles inconsistency");
205 job.traces_filenames.resize(job.app_size);
207 XBT_INFO("Job read: app='%s', file='%s', size=%d, start=%d, "
209 job.smpi_app_name.c_str(), filename_unprefixed.c_str(), job.app_size, job.starting_time,
211 jobs.push_back(new Job(std::move(job)));
212 } catch (const std::invalid_argument& e) {
213 xbt_die("Bad line '%s' of file '%s': %s.\n", line.c_str(), workload_file.c_str(), e.what());
218 // Jobs are sorted by ascending date, then by lexicographical order of their
220 sort(jobs.begin(), jobs.end(), job_comparator);
222 for (unsigned int i = 0; i < jobs.size(); ++i)
223 jobs[i]->unique_job_number = i;
228 int main(int argc, char* argv[])
231 "Usage: %s platform_file workload_file initial_noise noise_between_jobs\n"
232 "\tExample: %s platform.xml workload_compute\n",
235 // Simulation setting
236 simgrid::s4u::Engine e(&argc, argv);
237 e.load_platform(argv[1]);
238 hosts = e.get_all_hosts();
239 xbt_assert(hosts.size() >= 4, "The given platform should contain at least 4 hosts (found %zu).", hosts.size());
241 // Let's retrieve all SMPI jobs
242 std::vector<Job*> jobs = all_jobs(argv[2]);
244 // Let's register them
245 for (const Job* job : jobs)
246 SMPI_app_instance_register(job->smpi_app_name.c_str(), nullptr, job->app_size);
250 // Read noise arguments
251 int initial_noise = std::stoi(argv[3]);
252 xbt_assert(initial_noise >= 0, "Invalid initial_noise argument");
254 noise_between_jobs = std::stoi(argv[4]);
255 xbt_assert(noise_between_jobs >= 0, "Invalid noise_between_jobs argument");
257 if (initial_noise > 0) {
258 XBT_DEBUG("Popping %d noise processes", initial_noise);
259 pop_some_processes(initial_noise, hosts[0]);
262 // Let's execute the workload
263 simgrid::s4u::Actor::create("workload_executor", hosts[0], workload_executor_process, &jobs);
266 XBT_INFO("Simulation finished! Final time: %g", e.get_clock());
270 for (const Job* job : jobs)