1 /* Copyright (c) 2007-2021. 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 "xbt/config.hpp"
17 #include "xbt/file.hpp"
19 #include "src/smpi/include/smpi_actor.hpp"
20 #include <unordered_map>
32 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)");
34 static simgrid::config::Flag<double>
35 smpi_wtime_sleep("smpi/wtime",
36 "Minimum time to inject inside a call to MPI_Wtime(), gettimeofday() and clock_gettime()",
37 1e-8 /* Documented to be 10 ns */);
39 // Private execute_flops used by smpi_execute and smpi_execute_benched
40 void private_execute_flops(double flops) {
41 xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
42 XBT_DEBUG("Handle real computation time: %f flops", flops);
43 simgrid::s4u::this_actor::exec_init(flops)
44 ->set_name("computation")
45 ->set_tracing_category(smpi_process()->get_tracing_category())
48 smpi_switch_data_segment(simgrid::s4u::Actor::self());
51 void smpi_execute_flops(double flops)
53 private_execute_flops(flops);
56 void smpi_execute(double duration)
58 if (duration >= smpi_cfg_cpu_thresh()) {
59 XBT_DEBUG("Sleep for %g to handle real computation time", duration);
60 private_execute_flops(duration * smpi_cfg_host_speed());
62 XBT_DEBUG("Real computation took %g while option smpi/cpu-threshold is set to %g => ignore it", duration,
63 smpi_cfg_cpu_thresh());
67 void smpi_execute_benched(double duration)
70 double speed = sg_host_get_speed(sg_host_self());
71 smpi_execute_flops(duration*speed);
75 void smpi_execute_flops_benched(double flops) {
77 smpi_execute_flops(flops);
81 void smpi_bench_begin()
83 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
84 smpi_switch_data_segment(simgrid::s4u::Actor::self());
87 if (MC_is_active() || MC_record_replay_is_active())
91 if (not smpi_cfg_papi_events_file().empty()) {
92 int event_set = smpi_process()->papi_event_set();
93 // PAPI_start sets everything to 0! See man(3) PAPI_start
94 if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized() && event_set)
95 xbt_assert(PAPI_start(event_set) == PAPI_OK,
96 "Could not start PAPI counters (TODO: this needs some proper handling).");
99 xbt_os_threadtimer_start(smpi_process()->timer());
102 double smpi_adjust_comp_speed(){
104 if (smpi_cfg_comp_adjustment_file()[0] != '\0') {
105 const smpi_trace_call_location_t* loc = smpi_process()->call_location();
106 std::string key = loc->get_composed_key();
107 std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
108 if (it != location2speedup.end()) {
109 speedup = it->second;
115 void smpi_bench_end()
117 if (MC_is_active() || MC_record_replay_is_active())
120 xbt_os_timer_t timer = smpi_process()->timer();
121 xbt_os_threadtimer_stop(timer);
125 * An MPI function has been called and now is the right time to update
126 * our PAPI counters for this process.
128 if (not smpi_cfg_papi_events_file().empty()) {
129 papi_counter_t& counter_data = smpi_process()->papi_counters();
130 int event_set = smpi_process()->papi_event_set();
131 std::vector<long long> event_values(counter_data.size());
134 xbt_assert(PAPI_stop(event_set, &event_values[0]) == PAPI_OK, "Could not stop PAPI counters.");
135 for (unsigned int i = 0; i < counter_data.size(); i++)
136 counter_data[i].second += event_values[i];
140 if (smpi_process()->sampling()) {
141 XBT_CRITICAL("Cannot do recursive benchmarks.");
142 XBT_CRITICAL("Are you trying to make a call to MPI within an SMPI_SAMPLE_ block?");
143 xbt_backtrace_display_current();
144 xbt_die("Aborting.");
147 // Maybe we need to artificially speed up or slow down our computation based on our statistical analysis.
148 // Simulate the benchmarked computation unless disabled via command-line argument
149 if (smpi_cfg_simulate_computation()) {
150 smpi_execute(xbt_os_timer_elapsed(timer)/smpi_adjust_comp_speed());
154 if (not smpi_cfg_papi_events_file().empty() && TRACE_smpi_is_enabled()) {
155 simgrid::instr::Container* container =
156 simgrid::instr::Container::by_name(std::string("rank-") + std::to_string(simgrid::s4u::this_actor::get_pid()));
157 const papi_counter_t& counter_data = smpi_process()->papi_counters();
159 for (auto const& pair : counter_data) {
160 container->get_variable(pair.first)->set_event(simgrid::s4u::Engine::get_clock(), pair.second);
165 simgrid::smpi::utils::add_benched_time(xbt_os_timer_elapsed(timer));
168 /* Private sleep function used by smpi_sleep(), smpi_usleep() and friends */
169 static unsigned int private_sleep(double secs)
173 XBT_DEBUG("Sleep for: %lf secs", secs);
174 aid_t pid = simgrid::s4u::this_actor::get_pid();
175 TRACE_smpi_sleeping_in(pid, secs);
177 simgrid::s4u::this_actor::sleep_for(secs);
179 TRACE_smpi_sleeping_out(pid);
185 unsigned int smpi_sleep(unsigned int secs)
187 if (not smpi_process())
189 return private_sleep(secs);
192 int smpi_usleep(useconds_t usecs)
194 if (not smpi_process())
195 return usleep(usecs);
196 return static_cast<int>(private_sleep(usecs / 1000000.0));
199 #if _POSIX_TIMERS > 0
200 int smpi_nanosleep(const struct timespec* tp, struct timespec* t)
202 if (not smpi_process())
203 return nanosleep(tp,t);
204 return static_cast<int>(private_sleep(tp->tv_sec + tp->tv_nsec / 1000000000.0));
208 int smpi_gettimeofday(struct timeval* tv, struct timezone* tz)
210 if (not smpi_process()->initialized() || smpi_process()->finalized() || smpi_process()->sampling())
211 return gettimeofday(tv, tz);
214 double now = simgrid::s4u::Engine::get_clock();
216 tv->tv_sec = static_cast<time_t>(now);
218 tv->tv_usec = static_cast<useconds_t>((now - tv->tv_sec) * 1e6);
220 tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
223 if (smpi_wtime_sleep > 0)
224 simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
229 #if _POSIX_TIMERS > 0
230 int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp)
236 if (not smpi_process()->initialized() || smpi_process()->finalized() || smpi_process()->sampling())
237 return clock_gettime(clk_id, tp);
238 //there is only one time in SMPI, so clk_id is ignored.
240 double now = simgrid::s4u::Engine::get_clock();
241 tp->tv_sec = static_cast<time_t>(now);
242 tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
243 if (smpi_wtime_sleep > 0)
244 simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
250 double smpi_mpi_wtime()
253 if (smpi_process()->initialized() && not smpi_process()->finalized() && not smpi_process()->sampling()) {
255 time = simgrid::s4u::Engine::get_clock();
256 if (smpi_wtime_sleep > 0)
257 simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
260 time = simgrid::s4u::Engine::get_clock();
265 extern double sg_surf_precision;
266 unsigned long long smpi_rastro_resolution ()
269 double resolution = (1/sg_surf_precision);
271 return static_cast<unsigned long long>(resolution);
274 unsigned long long smpi_rastro_timestamp ()
277 double now = simgrid::s4u::Engine::get_clock();
279 auto sec = static_cast<unsigned long long>(now);
280 unsigned long long pre = (now - sec) * smpi_rastro_resolution();
282 return sec * smpi_rastro_resolution() + pre;
285 /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
287 class SampleLocation : public std::string {
289 SampleLocation(bool global, const char* file, int line) : std::string(std::string(file) + ":" + std::to_string(line))
292 this->append(":" + std::to_string(simgrid::s4u::this_actor::get_pid()));
298 double threshold; /* maximal stderr requested (if positive) */
299 double relstderr; /* observed stderr so far */
300 double mean; /* mean of benched times, to be used if the block is disabled */
301 double sum; /* sum of benched times (to compute the mean and stderr) */
302 double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
303 int iters; /* amount of requested iterations */
304 int count; /* amount of iterations done so far */
305 bool benching; /* true: we are benchmarking; false: we have enough data, no bench anymore */
307 bool need_more_benchs() const;
310 bool LocalData::need_more_benchs() const
312 bool res = (count < iters) || (threshold > 0.0 && (count < 2 || // not enough data
313 relstderr > threshold // stderr too high yet
315 XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
316 (res ? "need more data" : "enough benchs"), count, iters, relstderr, threshold, mean);
320 std::unordered_map<SampleLocation, LocalData, std::hash<std::string>> samples;
323 void smpi_sample_1(int global, const char *file, int line, int iters, double threshold)
325 SampleLocation loc(global, file, line);
326 if (not smpi_process()->sampling()) { /* Only at first call when benchmarking, skip for next ones */
327 smpi_bench_end(); /* Take time from previous, unrelated computation into account */
328 smpi_process()->set_sampling(1);
331 auto insert = samples.emplace(loc, LocalData{
332 threshold, // threshold
339 true // benching (if we have no data, we need at least one)
342 XBT_DEBUG("XXXXX First time ever on benched nest %s.", loc.c_str());
343 xbt_assert(threshold > 0 || iters > 0,
344 "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
346 LocalData& data = insert.first->second;
347 if (data.iters != iters || data.threshold != threshold) {
348 XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
349 "How did you manage to give two numbers at the same line??",
350 loc.c_str(), data.iters, data.threshold, iters, threshold);
354 // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate
355 // the computation instead
356 data.benching = data.need_more_benchs();
357 XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc.c_str(),
358 (data.benching ? "more benching needed" : "we have enough data, skip computes"));
362 int smpi_sample_2(int global, const char *file, int line, int iter_count)
364 SampleLocation loc(global, file, line);
366 XBT_DEBUG("sample2 %s %d", loc.c_str(), iter_count);
367 auto sample = samples.find(loc);
368 xbt_assert(sample != samples.end(),
369 "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
370 const LocalData& data = sample->second;
373 // we need to run a new bench
374 XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f; total:%f",
375 data.count, data.iters, data.relstderr, data.threshold, data.mean, data.sum);
378 // Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
379 //ran one bench and need to bail out now that our job is done). Just sleep instead
380 if (not data.need_more_benchs()){
381 XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f). "
382 "Mean is %f, will be injected %d times",
383 data.count, data.iters, data.relstderr, data.threshold, data.mean, iter_count);
385 //we ended benchmarking, let's inject all the time, now, and fast forward out of the loop.
386 smpi_process()->set_sampling(0);
387 smpi_execute(data.mean*iter_count);
391 XBT_DEBUG("Skipping - Benchmark already performed - accumulating time");
392 xbt_os_threadtimer_start(smpi_process()->timer());
398 void smpi_sample_3(int global, const char *file, int line)
400 SampleLocation loc(global, file, line);
402 XBT_DEBUG("sample3 %s", loc.c_str());
403 auto sample = samples.find(loc);
404 xbt_assert(sample != samples.end(),
405 "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
406 LocalData& data = sample->second;
408 if (not data.benching)
411 // ok, benchmarking this loop is over
412 xbt_os_threadtimer_stop(smpi_process()->timer());
416 double period = xbt_os_timer_elapsed(smpi_process()->timer());
418 data.sum_pow2 += period * period;
419 double n = data.count;
420 data.mean = data.sum / n;
421 data.relstderr = sqrt((data.sum_pow2 / n - data.mean * data.mean) / n) / data.mean;
423 XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)",
424 data.count, data.mean, data.relstderr, period);
426 // That's enough for now, prevent sample_2 to run the same code over and over
427 data.benching = false;
430 int smpi_sample_exit(int global, const char *file, int line, int iter_count){
431 if (smpi_process()->sampling()){
432 SampleLocation loc(global, file, line);
434 XBT_DEBUG("sample exit %s", loc.c_str());
435 auto sample = samples.find(loc);
436 xbt_assert(sample != samples.end(),
437 "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
439 if (smpi_process()->sampling()){//end of loop, but still sampling needed
440 const LocalData& data = sample->second;
441 smpi_process()->set_sampling(0);
442 smpi_execute(data.mean * iter_count);
449 smpi_trace_call_location_t* smpi_trace_get_call_location()
451 return smpi_process()->call_location();
454 void smpi_trace_set_call_location(const char* file, const int line)
456 smpi_trace_call_location_t* loc = smpi_process()->call_location();
458 loc->previous_filename = loc->filename;
459 loc->previous_linenumber = loc->linenumber;
460 if(not smpi_cfg_trace_call_use_absolute_path())
461 loc->filename = simgrid::xbt::Path(file).get_base_name();
463 loc->filename = file;
464 loc->linenumber = line;
467 /** Required for Fortran bindings */
468 void smpi_trace_set_call_location_(const char* file, const int* line)
470 smpi_trace_set_call_location(file, *line);
473 /** Required for Fortran if -fsecond-underscore is activated */
474 void smpi_trace_set_call_location__(const char* file, const int* line)
476 smpi_trace_set_call_location(file, *line);
479 void smpi_bench_destroy()
484 int smpi_getopt_long_only (int argc, char *const *argv, const char *options,
485 const struct option * long_options, int *opt_index)
488 optind = smpi_process()->get_optind();
489 int ret = getopt_long_only (argc, argv, options, long_options, opt_index);
491 smpi_process()->set_optind(optind);
495 int smpi_getopt_long (int argc, char *const *argv, const char *options,
496 const struct option * long_options, int *opt_index)
499 optind = smpi_process()->get_optind();
500 int ret = getopt_long (argc, argv, options, long_options, opt_index);
502 smpi_process()->set_optind(optind);
506 int smpi_getopt (int argc, char *const *argv, const char *options)
509 optind = smpi_process()->get_optind();
510 int ret = getopt (argc, argv, options);
512 smpi_process()->set_optind(optind);