1 /* Copyright (c) 2007-2022. 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. */
8 #include "simgrid/host.h"
9 #include "simgrid/modelchecker.h"
10 #include "simgrid/s4u/Engine.hpp"
11 #include "simgrid/s4u/Exec.hpp"
12 #include "smpi_comm.hpp"
13 #include "smpi_utils.hpp"
14 #include "src/internal_config.h"
15 #include "src/mc/mc_replay.hpp"
16 #include "src/surf/surf_interface.hpp" // sg_surf_precision
17 #include "xbt/config.hpp"
18 #include "xbt/file.hpp"
20 #include "src/smpi/include/smpi_actor.hpp"
21 #include <unordered_map>
33 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)");
35 static simgrid::config::Flag<double>
36 smpi_wtime_sleep("smpi/wtime",
37 "Minimum time to inject inside a call to MPI_Wtime(), gettimeofday() and clock_gettime()",
38 1e-8 /* Documented to be 10 ns */);
40 // Private execute_flops used by smpi_execute and smpi_execute_benched
41 void private_execute_flops(double flops) {
42 xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
43 XBT_DEBUG("Handle real computation time: %f flops", flops);
44 simgrid::s4u::this_actor::exec_init(flops)
45 ->set_name("computation")
46 ->set_tracing_category(smpi_process()->get_tracing_category())
49 smpi_switch_data_segment(simgrid::s4u::Actor::self());
52 void smpi_execute_flops(double flops)
54 private_execute_flops(flops);
57 void smpi_execute(double duration)
59 if (duration >= smpi_cfg_cpu_thresh()) {
60 XBT_DEBUG("Sleep for %gs (host time) to handle real computation time", duration);
61 private_execute_flops(duration * smpi_cfg_host_speed());
63 XBT_DEBUG("Real computation took %g while option smpi/cpu-threshold is set to %g => ignore it", duration,
64 smpi_cfg_cpu_thresh());
68 void smpi_execute_benched(double duration)
70 const SmpiBenchGuard suspend_bench;
71 double speed = sg_host_get_speed(sg_host_self());
72 smpi_execute_flops(duration*speed);
75 void smpi_execute_flops_benched(double flops) {
76 const SmpiBenchGuard suspend_bench;
77 smpi_execute_flops(flops);
80 void smpi_bench_begin()
82 smpi_switch_data_segment(simgrid::s4u::Actor::self());
84 if (MC_is_active() || MC_record_replay_is_active())
88 if (not smpi_cfg_papi_events_file().empty()) {
89 int event_set = smpi_process()->papi_event_set();
90 // PAPI_start sets everything to 0! See man(3) PAPI_start
91 if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized() && event_set)
92 xbt_assert(PAPI_start(event_set) == PAPI_OK,
93 "Could not start PAPI counters (TODO: this needs some proper handling).");
96 xbt_os_threadtimer_start(smpi_process()->timer());
99 double smpi_adjust_comp_speed(){
101 if (smpi_cfg_comp_adjustment_file()[0] != '\0') {
102 const smpi_trace_call_location_t* loc = smpi_process()->call_location();
103 std::string key = loc->get_composed_key();
104 std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
105 if (it != location2speedup.end()) {
106 speedup = it->second;
112 void smpi_bench_end()
114 if (MC_is_active() || MC_record_replay_is_active())
117 xbt_os_timer_t timer = smpi_process()->timer();
118 xbt_os_threadtimer_stop(timer);
122 * An MPI function has been called and now is the right time to update
123 * our PAPI counters for this process.
125 if (not smpi_cfg_papi_events_file().empty()) {
126 papi_counter_t& counter_data = smpi_process()->papi_counters();
127 int event_set = smpi_process()->papi_event_set();
128 std::vector<long long> event_values(counter_data.size());
131 xbt_assert(PAPI_stop(event_set, &event_values[0]) == PAPI_OK, "Could not stop PAPI counters.");
132 for (unsigned int i = 0; i < counter_data.size(); i++)
133 counter_data[i].second += event_values[i];
137 if (smpi_process()->sampling()) {
138 XBT_CRITICAL("Cannot do recursive benchmarks.");
139 XBT_CRITICAL("Are you trying to make a call to MPI within an SMPI_SAMPLE_ block?");
140 xbt_backtrace_display_current();
141 xbt_die("Aborting.");
144 // Maybe we need to artificially speed up or slow down our computation based on our statistical analysis.
145 // Simulate the benchmarked computation unless disabled via command-line argument
146 if (smpi_cfg_simulate_computation()) {
147 smpi_execute(xbt_os_timer_elapsed(timer)/smpi_adjust_comp_speed());
151 if (not smpi_cfg_papi_events_file().empty() && TRACE_smpi_is_enabled()) {
152 simgrid::instr::Container* container =
153 simgrid::instr::Container::by_name(std::string("rank-") + std::to_string(simgrid::s4u::this_actor::get_pid()));
154 const papi_counter_t& counter_data = smpi_process()->papi_counters();
156 for (auto const& [counter, value] : counter_data) {
157 container->get_variable(counter)->set_event(simgrid::s4u::Engine::get_clock(), static_cast<double>(value));
162 simgrid::smpi::utils::add_benched_time(xbt_os_timer_elapsed(timer));
165 /* Private sleep function used by smpi_sleep(), smpi_usleep() and friends */
166 static void private_sleep(double secs)
168 const SmpiBenchGuard suspend_bench;
170 XBT_DEBUG("Sleep for: %lf secs", secs);
171 aid_t pid = simgrid::s4u::this_actor::get_pid();
172 TRACE_smpi_sleeping_in(pid, secs);
173 simgrid::s4u::this_actor::sleep_for(secs);
174 TRACE_smpi_sleeping_out(pid);
177 unsigned int smpi_sleep(unsigned int secs)
179 if (not smpi_process())
185 int smpi_usleep(useconds_t usecs)
187 if (not smpi_process())
188 return usleep(usecs);
189 private_sleep(static_cast<double>(usecs) / 1e6);
193 #if _POSIX_TIMERS > 0
194 int smpi_nanosleep(const struct timespec* tp, struct timespec* t)
196 if (not smpi_process())
197 return nanosleep(tp,t);
198 private_sleep(static_cast<double>(tp->tv_sec) + static_cast<double>(tp->tv_nsec) / 1e9);
203 int smpi_gettimeofday(struct timeval* tv, struct timezone* tz)
205 if (not smpi_process()->initialized() || smpi_process()->finalized() || smpi_process()->sampling())
206 return gettimeofday(tv, tz);
208 const SmpiBenchGuard suspend_bench;
210 double now = simgrid::s4u::Engine::get_clock();
211 double secs = trunc(now);
212 double usecs = (now - secs) * 1e6;
213 tv->tv_sec = static_cast<time_t>(secs);
214 tv->tv_usec = static_cast<decltype(tv->tv_usec)>(usecs); // suseconds_t (or useconds_t on WIN32)
216 if (smpi_wtime_sleep > 0)
217 simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
221 #if _POSIX_TIMERS > 0
222 int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp)
228 if (not smpi_process()->initialized() || smpi_process()->finalized() || smpi_process()->sampling())
229 return clock_gettime(clk_id, tp);
230 //there is only one time in SMPI, so clk_id is ignored.
231 const SmpiBenchGuard suspend_bench;
232 double now = simgrid::s4u::Engine::get_clock();
233 double secs = trunc(now);
234 double nsecs = (now - secs) * 1e9;
235 tp->tv_sec = static_cast<time_t>(secs);
236 tp->tv_nsec = static_cast<long int>(nsecs);
237 if (smpi_wtime_sleep > 0)
238 simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
243 double smpi_mpi_wtime()
246 if (smpi_process()->initialized() && not smpi_process()->finalized() && not smpi_process()->sampling()) {
247 const SmpiBenchGuard suspend_bench;
248 time = simgrid::s4u::Engine::get_clock();
249 if (smpi_wtime_sleep > 0)
250 simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
252 time = simgrid::s4u::Engine::get_clock();
257 // Used by Akypuera (https://github.com/schnorr/akypuera)
258 unsigned long long smpi_rastro_resolution ()
260 const SmpiBenchGuard suspend_bench;
261 return static_cast<unsigned long long>(1.0 / sg_surf_precision);
264 unsigned long long smpi_rastro_timestamp ()
266 const SmpiBenchGuard suspend_bench;
267 return static_cast<unsigned long long>(simgrid::s4u::Engine::get_clock() / sg_surf_precision);
270 /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
272 class SampleLocation : public std::string {
274 SampleLocation(bool global, const char* file, const char* tag) : std::string(std::string(file) + ":" + std::string(tag))
277 this->append(":" + std::to_string(simgrid::s4u::this_actor::get_pid()));
283 double threshold; /* maximal stderr requested (if positive) */
284 double relstderr; /* observed stderr so far */
285 double mean; /* mean of benched times, to be used if the block is disabled */
286 double sum; /* sum of benched times (to compute the mean and stderr) */
287 double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
288 int iters; /* amount of requested iterations */
289 int count; /* amount of iterations done so far */
290 bool benching; /* true: we are benchmarking; false: we have enough data, no bench anymore */
292 bool need_more_benchs() const;
295 bool LocalData::need_more_benchs() const
297 bool res = (count < iters) && (threshold < 0.0 || count < 2 || // not enough data
298 relstderr >= threshold); // stderr too high yet
299 XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
300 (res ? "need more data" : "enough benchs"), count, iters, relstderr, threshold, mean);
304 std::unordered_map<SampleLocation, LocalData, std::hash<std::string>> samples;
307 int smpi_sample_cond(int global, const char* file, const char* tag, int iters, double threshold, int iter_count)
309 SampleLocation loc(global, file, tag);
310 if (not smpi_process()->sampling()) { /* Only at first call when benchmarking, skip for next ones */
311 smpi_bench_end(); /* Take time from previous, unrelated computation into account */
312 smpi_process()->set_sampling(1);
315 auto [sample, inserted] = samples.try_emplace(loc,
317 threshold, // threshold
324 true // benching (if we have no data, we need at least one)
326 LocalData& data = sample->second;
329 XBT_DEBUG("XXXXX First time ever on benched nest %s.", loc.c_str());
330 xbt_assert(threshold > 0 || iters > 0,
331 "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
333 if (data.iters != iters || data.threshold != threshold) {
334 XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
335 "How did you manage to give two numbers at the same line??",
336 loc.c_str(), data.iters, data.threshold, iters, threshold);
340 // if we already have some data, check whether we should get one more bench or whether we should emulate
341 // the computation instead
342 data.benching = data.need_more_benchs();
343 XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc.c_str(),
344 (data.benching ? "more benching needed" : "we have enough data, skip computes"));
347 XBT_DEBUG("sample cond %s %d", loc.c_str(), iter_count);
349 // we need to run a new bench
350 XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f; total:%f",
351 data.count, data.iters, data.relstderr, data.threshold, data.mean, data.sum);
354 // Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
355 //ran one bench and need to bail out now that our job is done). Just sleep instead
356 if (not data.need_more_benchs()){
357 XBT_DEBUG("No benchmark (either no need, or just ran one): count (%d) >= iter (%d) (or <2) or stderr (%f) < thres (%f), or thresh is negative and ignored. "
358 "Mean is %f, will be injected %d times",
359 data.count, data.iters, data.relstderr, data.threshold, data.mean, iter_count);
361 //we ended benchmarking, let's inject all the time, now, and fast forward out of the loop.
362 smpi_process()->set_sampling(0);
363 smpi_execute(data.mean*iter_count);
367 XBT_DEBUG("Skipping - Benchmark already performed - accumulating time");
368 xbt_os_threadtimer_start(smpi_process()->timer());
374 void smpi_sample_iter(int global, const char* file, const char* tag)
376 SampleLocation loc(global, file, tag);
378 XBT_DEBUG("sample iter %s", loc.c_str());
379 auto sample = samples.find(loc);
380 xbt_assert(sample != samples.end(),
381 "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
382 LocalData& data = sample->second;
383 xbt_assert(data.benching);
385 // ok, benchmarking this loop is over
386 xbt_os_threadtimer_stop(smpi_process()->timer());
390 double period = xbt_os_timer_elapsed(smpi_process()->timer());
392 data.sum_pow2 += period * period;
393 double n = data.count;
394 data.mean = data.sum / n;
395 data.relstderr = sqrt((data.sum_pow2 / n - data.mean * data.mean) / n) / data.mean;
397 XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)",
398 data.count, data.mean, data.relstderr, period);
401 int smpi_sample_exit(int global, const char *file, const char* tag, int iter_count){
402 if (smpi_process()->sampling()){
403 SampleLocation loc(global, file, tag);
405 XBT_DEBUG("sample exit %s", loc.c_str());
406 auto sample = samples.find(loc);
407 xbt_assert(sample != samples.end(),
408 "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
410 if (smpi_process()->sampling()){//end of loop, but still sampling needed
411 const LocalData& data = sample->second;
412 smpi_process()->set_sampling(0);
413 smpi_execute(data.mean * iter_count);
420 smpi_trace_call_location_t* smpi_trace_get_call_location()
422 return smpi_process()->call_location();
425 void smpi_trace_set_call_location(const char* file, const int line)
427 smpi_trace_call_location_t* loc = smpi_process()->call_location();
429 loc->previous_filename = loc->filename;
430 loc->previous_linenumber = loc->linenumber;
431 if(not smpi_cfg_trace_call_use_absolute_path())
432 loc->filename = simgrid::xbt::Path(file).get_base_name();
434 loc->filename = file;
435 loc->linenumber = line;
438 /** Required for Fortran bindings */
439 void smpi_trace_set_call_location_(const char* file, const int* line)
441 smpi_trace_set_call_location(file, *line);
444 /** Required for Fortran if -fsecond-underscore is activated */
445 void smpi_trace_set_call_location__(const char* file, const int* line)
447 smpi_trace_set_call_location(file, *line);
450 void smpi_bench_destroy()
455 int smpi_getopt_long_only (int argc, char *const *argv, const char *options,
456 const struct option * long_options, int *opt_index)
459 optind = smpi_process()->get_optind();
460 int ret = getopt_long_only (argc, argv, options, long_options, opt_index);
462 smpi_process()->set_optind(optind);
466 int smpi_getopt_long (int argc, char *const *argv, const char *options,
467 const struct option * long_options, int *opt_index)
470 optind = smpi_process()->get_optind();
471 int ret = getopt_long (argc, argv, options, long_options, opt_index);
473 smpi_process()->set_optind(optind);
477 int smpi_getopt (int argc, char *const *argv, const char *options)
480 optind = smpi_process()->get_optind();
481 int ret = getopt (argc, argv, options);
483 smpi_process()->set_optind(optind);
488 return static_cast<pid_t>(simgrid::s4u::this_actor::get_pid());