1 /* Copyright (c) 2009-2019. 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 XBT_INFO("Replaying rank %d of job %d (smpi_app '%s')", rank, job->unique_job_number, job->smpi_app_name.c_str());
60 smpi_replay_run(job->smpi_app_name.c_str(), rank, 0, job->traces_filenames[rank].c_str());
61 XBT_INFO("Finished replaying rank %d of job %d (smpi_app '%s')", rank, job->unique_job_number,
62 job->smpi_app_name.c_str());
67 // Sleeps for a given amount of time
68 static int sleeper_process(int* param)
70 XBT_DEBUG("Sleeping for %d seconds", *param);
71 simgrid::s4u::this_actor::sleep_for(*param);
78 // Launches some sleeper processes
79 static void pop_some_processes(int nb_processes, simgrid::s4u::Host* host)
81 for (int i = 0; i < nb_processes; ++i) {
84 simgrid::s4u::Actor::create("meh", host, sleeper_process, param);
88 static int job_executor_process(Job* job)
90 XBT_INFO("Executing job %d (smpi_app '%s')", job->unique_job_number, job->smpi_app_name.c_str());
92 simgrid::s4u::BarrierPtr barrier = simgrid::s4u::Barrier::create(job->app_size + 1);
94 for (int i = 0; i < job->app_size; ++i) {
95 char* str_pname = bprintf("rank_%d_%d", job->unique_job_number, i);
96 simgrid::s4u::Actor::create(str_pname, hosts[job->allocation[i]], smpi_replay_process, job, barrier, i);
102 simgrid::s4u::this_actor::sleep_for(1);
103 XBT_INFO("Finished job %d (smpi_app '%s')", job->unique_job_number, job->smpi_app_name.c_str());
108 // Executes a workload of SMPI processes
109 static int workload_executor_process(std::vector<Job*>* workload)
111 for (Job* job : *workload) {
112 // Let's wait until the job's waiting time if needed
113 double curr_time = simgrid::s4u::Engine::get_clock();
114 if (job->starting_time > curr_time) {
115 double time_to_sleep = (double)job->starting_time - curr_time;
116 XBT_INFO("Sleeping %g seconds (waiting for job %d, app '%s')", time_to_sleep, job->starting_time,
117 job->smpi_app_name.c_str());
118 simgrid::s4u::this_actor::sleep_for(time_to_sleep);
121 if (noise_between_jobs > 0) {
122 // Let's add some process noise
123 XBT_DEBUG("Popping %d noise processes before running job %d (app '%s')", noise_between_jobs,
124 job->unique_job_number, job->smpi_app_name.c_str());
125 pop_some_processes(noise_between_jobs, hosts[job->allocation[0]]);
128 // Let's finally run the job executor
129 char* str_pname = bprintf("job_%04d", job->unique_job_number);
130 XBT_INFO("Launching the job executor of job %d (app '%s')", job->unique_job_number, job->smpi_app_name.c_str());
131 simgrid::s4u::Actor::create(str_pname, hosts[job->allocation[0]], job_executor_process, job);
138 // Reads jobs from a workload file and returns them
139 static std::vector<Job*> all_jobs(const std::string& workload_file)
141 std::ifstream f(workload_file);
142 xbt_assert(f.is_open(), "Cannot open file '%s'.", workload_file.c_str());
143 std::vector<Job*> jobs;
145 simgrid::xbt::Path path(workload_file);
146 std::string dir = path.get_dir_name();
149 while (std::getline(f, line)) {
150 std::string app_name;
151 std::string filename_unprefixed;
156 std::istringstream is(line);
157 if (is >> app_name >> filename_unprefixed >> app_size >> starting_time >> alloc) {
160 job.smpi_app_name = app_name;
161 job.filename = dir + "/" + filename_unprefixed;
162 job.app_size = app_size;
163 job.starting_time = starting_time;
165 std::vector<std::string> subparts;
166 boost::split(subparts, alloc, boost::is_any_of(","), boost::token_compress_on);
168 if ((int)subparts.size() != job.app_size)
169 throw std::invalid_argument("size/alloc inconsistency");
171 job.allocation.resize(subparts.size());
172 for (unsigned int i = 0; i < subparts.size(); ++i)
173 job.allocation[i] = stoi(subparts[i]);
175 // Let's read the filename
176 std::ifstream traces_file(job.filename);
177 if (!traces_file.is_open())
178 throw std::invalid_argument("Cannot open file " + job.filename);
180 std::string traces_line;
181 while (std::getline(traces_file, traces_line)) {
182 boost::trim_right(traces_line);
183 job.traces_filenames.push_back(dir + "/" + traces_line);
186 if (static_cast<int>(job.traces_filenames.size()) < job.app_size)
187 throw std::invalid_argument("size/tracefiles inconsistency");
188 job.traces_filenames.resize(job.app_size);
190 XBT_INFO("Job read: app='%s', file='%s', size=%d, start=%d, "
192 job.smpi_app_name.c_str(), filename_unprefixed.c_str(), job.app_size, job.starting_time,
194 jobs.push_back(new Job(std::move(job)));
195 } catch (const std::invalid_argument& e) {
196 xbt_die("Bad line '%s' of file '%s': %s.\n", line.c_str(), workload_file.c_str(), e.what());
201 // Jobs are sorted by ascending date, then by lexicographical order of their
203 sort(jobs.begin(), jobs.end(), job_comparator);
205 for (unsigned int i = 0; i < jobs.size(); ++i)
206 jobs[i]->unique_job_number = i;
211 int main(int argc, char* argv[])
214 "Usage: %s platform_file workload_file initial_noise noise_between_jobs\n"
215 "\tExample: %s platform.xml workload_compute\n",
218 // Simulation setting
219 simgrid::s4u::Engine e(&argc, argv);
220 e.load_platform(argv[1]);
221 hosts = e.get_all_hosts();
222 xbt_assert(hosts.size() >= 4, "The given platform should contain at least 4 hosts (found %zu).", hosts.size());
224 // Let's retrieve all SMPI jobs
225 std::vector<Job*> jobs = all_jobs(argv[2]);
227 // Let's register them
228 for (const Job* job : jobs)
229 SMPI_app_instance_register(job->smpi_app_name.c_str(), nullptr, job->app_size);
233 // Read noise arguments
234 int initial_noise = std::stoi(argv[3]);
235 xbt_assert(initial_noise >= 0, "Invalid initial_noise argument");
237 noise_between_jobs = std::stoi(argv[4]);
238 xbt_assert(noise_between_jobs >= 0, "Invalid noise_between_jobs argument");
240 if (initial_noise > 0) {
241 XBT_DEBUG("Popping %d noise processes", initial_noise);
242 pop_some_processes(initial_noise, hosts[0]);
245 // Let's execute the workload
246 simgrid::s4u::Actor::create("workload", hosts[0], workload_executor_process, &jobs);
249 XBT_INFO("Simulation finished! Final time: %g", e.get_clock());
253 for (const Job* job : jobs)