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 %gs (host time) 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)
69 const SmpiBenchGuard suspend_bench;
70 double speed = sg_host_get_speed(sg_host_self());
71 smpi_execute_flops(duration*speed);
74 void smpi_execute_flops_benched(double flops) {
75 const SmpiBenchGuard suspend_bench;
76 smpi_execute_flops(flops);
79 void smpi_bench_begin()
81 smpi_switch_data_segment(simgrid::s4u::Actor::self());
83 if (MC_is_active() || MC_record_replay_is_active())
87 if (not smpi_cfg_papi_events_file().empty()) {
88 int event_set = smpi_process()->papi_event_set();
89 // PAPI_start sets everything to 0! See man(3) PAPI_start
90 if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized() && event_set)
91 xbt_assert(PAPI_start(event_set) == PAPI_OK,
92 "Could not start PAPI counters (TODO: this needs some proper handling).");
95 xbt_os_threadtimer_start(smpi_process()->timer());
98 double smpi_adjust_comp_speed(){
100 if (smpi_cfg_comp_adjustment_file()[0] != '\0') {
101 const smpi_trace_call_location_t* loc = smpi_process()->call_location();
102 std::string key = loc->get_composed_key();
103 std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
104 if (it != location2speedup.end()) {
105 speedup = it->second;
111 void smpi_bench_end()
113 if (MC_is_active() || MC_record_replay_is_active())
116 xbt_os_timer_t timer = smpi_process()->timer();
117 xbt_os_threadtimer_stop(timer);
121 * An MPI function has been called and now is the right time to update
122 * our PAPI counters for this process.
124 if (not smpi_cfg_papi_events_file().empty()) {
125 papi_counter_t& counter_data = smpi_process()->papi_counters();
126 int event_set = smpi_process()->papi_event_set();
127 std::vector<long long> event_values(counter_data.size());
130 xbt_assert(PAPI_stop(event_set, &event_values[0]) == PAPI_OK, "Could not stop PAPI counters.");
131 for (unsigned int i = 0; i < counter_data.size(); i++)
132 counter_data[i].second += event_values[i];
136 if (smpi_process()->sampling()) {
137 XBT_CRITICAL("Cannot do recursive benchmarks.");
138 XBT_CRITICAL("Are you trying to make a call to MPI within an SMPI_SAMPLE_ block?");
139 xbt_backtrace_display_current();
140 xbt_die("Aborting.");
143 // Maybe we need to artificially speed up or slow down our computation based on our statistical analysis.
144 // Simulate the benchmarked computation unless disabled via command-line argument
145 if (smpi_cfg_simulate_computation()) {
146 smpi_execute(xbt_os_timer_elapsed(timer)/smpi_adjust_comp_speed());
150 if (not smpi_cfg_papi_events_file().empty() && TRACE_smpi_is_enabled()) {
151 simgrid::instr::Container* container =
152 simgrid::instr::Container::by_name(std::string("rank-") + std::to_string(simgrid::s4u::this_actor::get_pid()));
153 const papi_counter_t& counter_data = smpi_process()->papi_counters();
155 for (auto const& pair : counter_data) {
156 container->get_variable(pair.first)->set_event(simgrid::s4u::Engine::get_clock(), pair.second);
161 simgrid::smpi::utils::add_benched_time(xbt_os_timer_elapsed(timer));
164 /* Private sleep function used by smpi_sleep(), smpi_usleep() and friends */
165 static void private_sleep(double secs)
167 const SmpiBenchGuard suspend_bench;
169 XBT_DEBUG("Sleep for: %lf secs", secs);
170 aid_t pid = simgrid::s4u::this_actor::get_pid();
171 TRACE_smpi_sleeping_in(pid, secs);
172 simgrid::s4u::this_actor::sleep_for(secs);
173 TRACE_smpi_sleeping_out(pid);
176 unsigned int smpi_sleep(unsigned int secs)
178 if (not smpi_process())
184 int smpi_usleep(useconds_t usecs)
186 if (not smpi_process())
187 return usleep(usecs);
188 private_sleep(static_cast<double>(usecs) / 1e6);
192 #if _POSIX_TIMERS > 0
193 int smpi_nanosleep(const struct timespec* tp, struct timespec* t)
195 if (not smpi_process())
196 return nanosleep(tp,t);
197 private_sleep(static_cast<double>(tp->tv_sec) + static_cast<double>(tp->tv_nsec) / 1e9);
202 int smpi_gettimeofday(struct timeval* tv, struct timezone* tz)
204 if (not smpi_process()->initialized() || smpi_process()->finalized() || smpi_process()->sampling())
205 return gettimeofday(tv, tz);
207 const SmpiBenchGuard suspend_bench;
209 double now = simgrid::s4u::Engine::get_clock();
210 double secs = trunc(now);
211 double usecs = (now - secs) * 1e6;
212 tv->tv_sec = static_cast<time_t>(secs);
213 tv->tv_usec = static_cast<decltype(tv->tv_usec)>(usecs); // suseconds_t (or useconds_t on WIN32)
215 if (smpi_wtime_sleep > 0)
216 simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
220 #if _POSIX_TIMERS > 0
221 int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp)
227 if (not smpi_process()->initialized() || smpi_process()->finalized() || smpi_process()->sampling())
228 return clock_gettime(clk_id, tp);
229 //there is only one time in SMPI, so clk_id is ignored.
230 const SmpiBenchGuard suspend_bench;
231 double now = simgrid::s4u::Engine::get_clock();
232 double secs = trunc(now);
233 double nsecs = (now - secs) * 1e9;
234 tp->tv_sec = static_cast<time_t>(secs);
235 tp->tv_nsec = static_cast<long int>(nsecs);
236 if (smpi_wtime_sleep > 0)
237 simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
242 double smpi_mpi_wtime()
245 if (smpi_process()->initialized() && not smpi_process()->finalized() && not smpi_process()->sampling()) {
246 const SmpiBenchGuard suspend_bench;
247 time = simgrid::s4u::Engine::get_clock();
248 if (smpi_wtime_sleep > 0)
249 simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
251 time = simgrid::s4u::Engine::get_clock();
256 extern double sg_surf_precision;
257 unsigned long long smpi_rastro_resolution ()
259 const SmpiBenchGuard suspend_bench;
260 double resolution = (1/sg_surf_precision);
261 return static_cast<unsigned long long>(resolution);
264 unsigned long long smpi_rastro_timestamp ()
266 const SmpiBenchGuard suspend_bench;
267 double now = simgrid::s4u::Engine::get_clock();
269 auto sec = static_cast<unsigned long long>(now);
270 unsigned long long pre = (now - sec) * smpi_rastro_resolution();
271 return sec * smpi_rastro_resolution() + pre;
274 /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
276 class SampleLocation : public std::string {
278 SampleLocation(bool global, const char* file, const char* tag) : std::string(std::string(file) + ":" + std::string(tag))
281 this->append(":" + std::to_string(simgrid::s4u::this_actor::get_pid()));
287 double threshold; /* maximal stderr requested (if positive) */
288 double relstderr; /* observed stderr so far */
289 double mean; /* mean of benched times, to be used if the block is disabled */
290 double sum; /* sum of benched times (to compute the mean and stderr) */
291 double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
292 int iters; /* amount of requested iterations */
293 int count; /* amount of iterations done so far */
294 bool benching; /* true: we are benchmarking; false: we have enough data, no bench anymore */
296 bool need_more_benchs() const;
299 bool LocalData::need_more_benchs() const
301 bool res = (count < iters) && (threshold < 0.0 || count < 2 || // not enough data
302 relstderr >= threshold); // stderr too high yet
303 XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
304 (res ? "need more data" : "enough benchs"), count, iters, relstderr, threshold, mean);
308 std::unordered_map<SampleLocation, LocalData, std::hash<std::string>> samples;
311 void smpi_sample_1(int global, const char *file, const char *tag, int iters, double threshold)
313 SampleLocation loc(global, file, tag);
314 if (not smpi_process()->sampling()) { /* Only at first call when benchmarking, skip for next ones */
315 smpi_bench_end(); /* Take time from previous, unrelated computation into account */
316 smpi_process()->set_sampling(1);
319 auto insert = samples.emplace(loc, LocalData{
320 threshold, // threshold
327 true // benching (if we have no data, we need at least one)
330 XBT_DEBUG("XXXXX First time ever on benched nest %s.", loc.c_str());
331 xbt_assert(threshold > 0 || iters > 0,
332 "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
334 LocalData& data = insert.first->second;
335 if (data.iters != iters || data.threshold != threshold) {
336 XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
337 "How did you manage to give two numbers at the same line??",
338 loc.c_str(), data.iters, data.threshold, iters, threshold);
342 // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate
343 // the computation instead
344 data.benching = data.need_more_benchs();
345 XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc.c_str(),
346 (data.benching ? "more benching needed" : "we have enough data, skip computes"));
350 int smpi_sample_2(int global, const char *file,const char *tag, int iter_count)
352 SampleLocation loc(global, file, tag);
354 XBT_DEBUG("sample2 %s %d", loc.c_str(), iter_count);
355 auto sample = samples.find(loc);
356 xbt_assert(sample != samples.end(),
357 "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
358 const LocalData& data = sample->second;
361 // we need to run a new bench
362 XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f; total:%f",
363 data.count, data.iters, data.relstderr, data.threshold, data.mean, data.sum);
366 // Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
367 //ran one bench and need to bail out now that our job is done). Just sleep instead
368 if (not data.need_more_benchs()){
369 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. "
370 "Mean is %f, will be injected %d times",
371 data.count, data.iters, data.relstderr, data.threshold, data.mean, iter_count);
373 //we ended benchmarking, let's inject all the time, now, and fast forward out of the loop.
374 smpi_process()->set_sampling(0);
375 smpi_execute(data.mean*iter_count);
379 XBT_DEBUG("Skipping - Benchmark already performed - accumulating time");
380 xbt_os_threadtimer_start(smpi_process()->timer());
386 void smpi_sample_3(int global, const char *file, const char* tag)
388 SampleLocation loc(global, file, tag);
390 XBT_DEBUG("sample3 %s", loc.c_str());
391 auto sample = samples.find(loc);
392 xbt_assert(sample != samples.end(),
393 "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
394 LocalData& data = sample->second;
396 if (not data.benching)
399 // ok, benchmarking this loop is over
400 xbt_os_threadtimer_stop(smpi_process()->timer());
404 double period = xbt_os_timer_elapsed(smpi_process()->timer());
406 data.sum_pow2 += period * period;
407 double n = data.count;
408 data.mean = data.sum / n;
409 data.relstderr = sqrt((data.sum_pow2 / n - data.mean * data.mean) / n) / data.mean;
411 XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)",
412 data.count, data.mean, data.relstderr, period);
414 // That's enough for now, prevent sample_2 to run the same code over and over
415 data.benching = false;
418 int smpi_sample_exit(int global, const char *file, const char* tag, int iter_count){
419 if (smpi_process()->sampling()){
420 SampleLocation loc(global, file, tag);
422 XBT_DEBUG("sample exit %s", loc.c_str());
423 auto sample = samples.find(loc);
424 xbt_assert(sample != samples.end(),
425 "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
427 if (smpi_process()->sampling()){//end of loop, but still sampling needed
428 const LocalData& data = sample->second;
429 smpi_process()->set_sampling(0);
430 smpi_execute(data.mean * iter_count);
437 smpi_trace_call_location_t* smpi_trace_get_call_location()
439 return smpi_process()->call_location();
442 void smpi_trace_set_call_location(const char* file, const int line)
444 smpi_trace_call_location_t* loc = smpi_process()->call_location();
446 loc->previous_filename = loc->filename;
447 loc->previous_linenumber = loc->linenumber;
448 if(not smpi_cfg_trace_call_use_absolute_path())
449 loc->filename = simgrid::xbt::Path(file).get_base_name();
451 loc->filename = file;
452 loc->linenumber = line;
455 /** Required for Fortran bindings */
456 void smpi_trace_set_call_location_(const char* file, const int* line)
458 smpi_trace_set_call_location(file, *line);
461 /** Required for Fortran if -fsecond-underscore is activated */
462 void smpi_trace_set_call_location__(const char* file, const int* line)
464 smpi_trace_set_call_location(file, *line);
467 void smpi_bench_destroy()
472 int smpi_getopt_long_only (int argc, char *const *argv, const char *options,
473 const struct option * long_options, int *opt_index)
476 optind = smpi_process()->get_optind();
477 int ret = getopt_long_only (argc, argv, options, long_options, opt_index);
479 smpi_process()->set_optind(optind);
483 int smpi_getopt_long (int argc, char *const *argv, const char *options,
484 const struct option * long_options, int *opt_index)
487 optind = smpi_process()->get_optind();
488 int ret = getopt_long (argc, argv, options, long_options, opt_index);
490 smpi_process()->set_optind(optind);
494 int smpi_getopt (int argc, char *const *argv, const char *options)
497 optind = smpi_process()->get_optind();
498 int ret = getopt (argc, argv, options);
500 smpi_process()->set_optind(optind);