1 /* Copyright (c) 2007-2019. 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/Exec.hpp"
11 #include "smpi_comm.hpp"
12 #include "src/internal_config.h"
13 #include "src/mc/mc_replay.hpp"
14 #include "xbt/config.hpp"
15 #include "xbt/file.hpp"
17 #include "src/smpi/include/smpi_actor.hpp"
18 #include <unordered_map>
29 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)");
31 static simgrid::config::Flag<double>
32 smpi_wtime_sleep("smpi/wtime",
33 "Minimum time to inject inside a call to MPI_Wtime(), gettimeofday() and clock_gettime()",
34 1e-8 /* Documented to be 10 ns */);
36 double smpi_total_benched_time = 0;
38 // Private execute_flops used by smpi_execute and smpi_execute_benched
39 void private_execute_flops(double flops) {
40 xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
41 XBT_DEBUG("Handle real computation time: %f flops", flops);
42 simgrid::s4u::this_actor::exec_init(flops)
43 ->set_name("computation")
44 ->set_tracing_category(smpi_process()->get_tracing_category())
47 smpi_switch_data_segment(simgrid::s4u::Actor::self());
50 void smpi_execute_flops(double flops) {
51 int rank = simgrid::s4u::this_actor::get_pid();
52 TRACE_smpi_computing_in(rank, flops);
54 private_execute_flops(flops);
56 TRACE_smpi_computing_out(rank);
59 void smpi_execute(double duration)
61 if (duration >= smpi_cfg_cpu_thresh()) {
62 XBT_DEBUG("Sleep for %g to handle real computation time", duration);
63 double flops = duration * smpi_cfg_host_speed();
64 int rank = simgrid::s4u::this_actor::get_pid();
65 TRACE_smpi_computing_in(rank, flops);
67 private_execute_flops(flops);
69 TRACE_smpi_computing_out(rank);
72 XBT_DEBUG("Real computation took %g while option smpi/cpu-threshold is set to %g => ignore it", duration,
73 smpi_cfg_cpu_thresh());
77 void smpi_execute_benched(double duration)
80 double speed = sg_host_speed(sg_host_self());
81 smpi_execute_flops(duration*speed);
85 void smpi_bench_begin()
87 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
88 smpi_switch_data_segment(simgrid::s4u::Actor::self());
91 if (MC_is_active() || MC_record_replay_is_active())
95 if (not smpi_cfg_papi_events_file().empty()) {
96 int event_set = smpi_process()->papi_event_set();
97 // PAPI_start sets everything to 0! See man(3) PAPI_start
98 if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized() && PAPI_start(event_set) != PAPI_OK) {
99 // TODO This needs some proper handling.
100 XBT_CRITICAL("Could not start PAPI counters.\n");
105 xbt_os_threadtimer_start(smpi_process()->timer());
108 double smpi_adjust_comp_speed(){
110 if (smpi_cfg_comp_adjustment_file()[0] != '\0') {
112 smpi_trace_call_location_t* loc = smpi_process()->call_location();
113 std::string key = loc->get_composed_key();
114 std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
115 if (it != location2speedup.end()) {
116 speedup = it->second;
122 void smpi_bench_end()
124 if (MC_is_active() || MC_record_replay_is_active())
127 xbt_os_timer_t timer = smpi_process()->timer();
128 xbt_os_threadtimer_stop(timer);
132 * An MPI function has been called and now is the right time to update
133 * our PAPI counters for this process.
135 if (not smpi_cfg_papi_events_file().empty()) {
136 papi_counter_t& counter_data = smpi_process()->papi_counters();
137 int event_set = smpi_process()->papi_event_set();
138 std::vector<long long> event_values = std::vector<long long>(counter_data.size());
140 if (PAPI_stop(event_set, &event_values[0]) != PAPI_OK) { // Error
141 XBT_CRITICAL("Could not stop PAPI counters.\n");
144 for (unsigned int i = 0; i < counter_data.size(); i++) {
145 counter_data[i].second += event_values[i];
151 if (smpi_process()->sampling()) {
152 XBT_CRITICAL("Cannot do recursive benchmarks.");
153 XBT_CRITICAL("Are you trying to make a call to MPI within a SMPI_SAMPLE_ block?");
154 xbt_backtrace_display_current();
155 xbt_die("Aborting.");
158 // Maybe we need to artificially speed up or slow down our computation based on our statistical analysis.
159 // Simulate the benchmarked computation unless disabled via command-line argument
160 if (smpi_cfg_simulate_computation()) {
161 smpi_execute(xbt_os_timer_elapsed(timer)/smpi_adjust_comp_speed());
165 if (not smpi_cfg_papi_events_file().empty() && TRACE_smpi_is_enabled()) {
166 container_t container =
167 simgrid::instr::Container::by_name(std::string("rank-") + std::to_string(simgrid::s4u::this_actor::get_pid()));
168 papi_counter_t& counter_data = smpi_process()->papi_counters();
170 for (auto const& pair : counter_data) {
171 simgrid::instr::VariableType* variable = static_cast<simgrid::instr::VariableType*>(container->type_->by_name(pair.first));
172 variable->set_event(SIMIX_get_clock(), pair.second);
177 smpi_total_benched_time += xbt_os_timer_elapsed(timer);
180 /* Private sleep function used by smpi_sleep(), smpi_usleep() and friends */
181 static unsigned int private_sleep(double secs)
185 XBT_DEBUG("Sleep for: %lf secs", secs);
186 int rank = simgrid::s4u::this_actor::get_pid();
187 TRACE_smpi_sleeping_in(rank, secs);
189 simgrid::s4u::this_actor::sleep_for(secs);
191 TRACE_smpi_sleeping_out(rank);
197 unsigned int smpi_sleep(unsigned int secs)
199 if (not smpi_process())
201 return private_sleep(secs);
204 int smpi_usleep(useconds_t usecs)
206 if (not smpi_process())
207 return usleep(usecs);
208 return static_cast<int>(private_sleep(usecs / 1000000.0));
211 #if _POSIX_TIMERS > 0
212 int smpi_nanosleep(const struct timespec* tp, struct timespec* t)
214 if (not smpi_process())
215 return nanosleep(tp,t);
216 return static_cast<int>(private_sleep(tp->tv_sec + tp->tv_nsec / 1000000000.0));
220 int smpi_gettimeofday(struct timeval* tv, struct timezone* tz)
222 if (not smpi_process())
223 return gettimeofday(tv, tz);
226 double now = SIMIX_get_clock();
228 tv->tv_sec = static_cast<time_t>(now);
230 tv->tv_usec = static_cast<useconds_t>((now - tv->tv_sec) * 1e6);
232 tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
235 if (smpi_wtime_sleep > 0)
236 simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
241 #if _POSIX_TIMERS > 0
242 int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp)
244 if (not smpi_process())
245 return clock_gettime(clk_id, tp);
246 //there is only one time in SMPI, so clk_id is ignored.
248 double now = SIMIX_get_clock();
250 tp->tv_sec = static_cast<time_t>(now);
251 tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
253 if (smpi_wtime_sleep > 0)
254 simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
260 double smpi_mpi_wtime()
263 if (smpi_process()->initialized() && not smpi_process()->finalized() && not smpi_process()->sampling()) {
265 time = SIMIX_get_clock();
266 if (smpi_wtime_sleep > 0)
267 simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
270 time = SIMIX_get_clock();
275 extern double sg_surf_precision;
276 unsigned long long smpi_rastro_resolution ()
279 double resolution = (1/sg_surf_precision);
281 return static_cast<unsigned long long>(resolution);
284 unsigned long long smpi_rastro_timestamp ()
287 double now = SIMIX_get_clock();
289 unsigned long long sec = static_cast<unsigned long long>(now);
290 unsigned long long pre = (now - sec) * smpi_rastro_resolution();
292 return static_cast<unsigned long long>(sec) * smpi_rastro_resolution() + pre;
295 /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
297 class SampleLocation : public std::string {
299 SampleLocation(bool global, const char* file, int line) : std::string(std::string(file) + ":" + std::to_string(line))
302 this->append(":" + std::to_string(simgrid::s4u::this_actor::get_pid()));
308 double threshold; /* maximal stderr requested (if positive) */
309 double relstderr; /* observed stderr so far */
310 double mean; /* mean of benched times, to be used if the block is disabled */
311 double sum; /* sum of benched times (to compute the mean and stderr) */
312 double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
313 int iters; /* amount of requested iterations */
314 int count; /* amount of iterations done so far */
315 bool benching; /* true: we are benchmarking; false: we have enough data, no bench anymore */
317 bool need_more_benchs() const;
320 bool LocalData::need_more_benchs() const
322 bool res = (count < iters) || (threshold > 0.0 && (count < 2 || // not enough data
323 relstderr > threshold // stderr too high yet
325 XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
326 (res ? "need more data" : "enough benchs"), count, iters, relstderr, threshold, mean);
330 std::unordered_map<SampleLocation, LocalData, std::hash<std::string>> samples;
333 void smpi_sample_1(int global, const char *file, int line, int iters, double threshold)
335 SampleLocation loc(global, file, line);
336 if (not smpi_process()->sampling()) { /* Only at first call when benchmarking, skip for next ones */
337 smpi_bench_end(); /* Take time from previous, unrelated computation into account */
338 smpi_process()->set_sampling(1);
341 auto insert = samples.emplace(loc, LocalData{
342 threshold, // threshold
349 true // benching (if we have no data, we need at least one)
352 XBT_DEBUG("XXXXX First time ever on benched nest %s.", loc.c_str());
353 xbt_assert(threshold > 0 || iters > 0,
354 "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
356 LocalData& data = insert.first->second;
357 if (data.iters != iters || data.threshold != threshold) {
358 XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
359 "How did you manage to give two numbers at the same line??",
360 loc.c_str(), data.iters, data.threshold, iters, threshold);
364 // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate
365 // the computation instead
366 data.benching = data.need_more_benchs();
367 XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc.c_str(),
368 (data.benching ? "more benching needed" : "we have enough data, skip computes"));
372 int smpi_sample_2(int global, const char *file, int line, int iter_count)
374 SampleLocation loc(global, file, line);
376 XBT_DEBUG("sample2 %s %d", loc.c_str(), iter_count);
377 auto sample = samples.find(loc);
378 if (sample == samples.end())
379 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
380 LocalData& data = sample->second;
383 // we need to run a new bench
384 XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f; total:%f",
385 data.count, data.iters, data.relstderr, data.threshold, data.mean, data.sum);
388 // Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
389 //ran one bench and need to bail out now that our job is done). Just sleep instead
390 if (not data.need_more_benchs()){
391 XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f). "
392 "Mean is %f, will be injected %d times",
393 data.count, data.iters, data.relstderr, data.threshold, data.mean, iter_count);
395 //we ended benchmarking, let's inject all the time, now, and fast forward out of the loop.
396 smpi_process()->set_sampling(0);
397 smpi_execute(data.mean*iter_count);
401 XBT_DEBUG("Skipping - Benchmark already performed - accumulating time");
402 xbt_os_threadtimer_start(smpi_process()->timer());
408 void smpi_sample_3(int global, const char *file, int line)
410 SampleLocation loc(global, file, line);
412 XBT_DEBUG("sample3 %s", loc.c_str());
413 auto sample = samples.find(loc);
414 if (sample == samples.end())
415 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
416 LocalData& data = sample->second;
418 if (not data.benching)
421 // ok, benchmarking this loop is over
422 xbt_os_threadtimer_stop(smpi_process()->timer());
426 double period = xbt_os_timer_elapsed(smpi_process()->timer());
428 data.sum_pow2 += period * period;
429 double n = data.count;
430 data.mean = data.sum / n;
431 data.relstderr = sqrt((data.sum_pow2 / n - data.mean * data.mean) / n) / data.mean;
433 XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)",
434 data.count, data.mean, data.relstderr, period);
436 // That's enough for now, prevent sample_2 to run the same code over and over
437 data.benching = false;
440 int smpi_sample_exit(int global, const char *file, int line, int iter_count){
441 if (smpi_process()->sampling()){
442 SampleLocation loc(global, file, line);
444 XBT_DEBUG("sample exit %s", loc.c_str());
445 auto sample = samples.find(loc);
446 if (sample == samples.end())
447 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
449 if (smpi_process()->sampling()){//end of loop, but still sampling needed
450 LocalData& data = sample->second;
451 smpi_process()->set_sampling(0);
452 smpi_execute(data.mean * iter_count);
459 smpi_trace_call_location_t* smpi_trace_get_call_location()
461 return smpi_process()->call_location();
464 void smpi_trace_set_call_location(const char* file, const int line)
466 smpi_trace_call_location_t* loc = smpi_process()->call_location();
468 loc->previous_filename = loc->filename;
469 loc->previous_linenumber = loc->linenumber;
470 if(not smpi_cfg_trace_call_use_absolute_path())
471 loc->filename = simgrid::xbt::Path(file).get_base_name();
473 loc->filename = file;
474 loc->linenumber = line;
477 /** Required for Fortran bindings */
478 void smpi_trace_set_call_location_(const char* file, int* line)
480 smpi_trace_set_call_location(file, *line);
483 /** Required for Fortran if -fsecond-underscore is activated */
484 void smpi_trace_set_call_location__(const char* file, int* line)
486 smpi_trace_set_call_location(file, *line);
489 void smpi_bench_destroy()
494 int smpi_getopt_long_only (int argc, char *const *argv, const char *options,
495 const struct option * long_options, int *opt_index)
498 optind = smpi_process()->get_optind();
499 int ret = getopt_long_only (argc, argv, options, long_options, opt_index);
501 smpi_process()->set_optind(optind);
505 int smpi_getopt_long (int argc, char *const *argv, const char *options,
506 const struct option * long_options, int *opt_index)
509 optind = smpi_process()->get_optind();
510 int ret = getopt_long (argc, argv, options, long_options, opt_index);
512 smpi_process()->set_optind(optind);
516 int smpi_getopt (int argc, char *const *argv, const char *options)
519 optind = smpi_process()->get_optind();
520 int ret = getopt (argc, argv, options);
522 smpi_process()->set_optind(optind);