1 /* Copyright (c) 2007-2018. 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 "src/simix/ActorImpl.hpp"
15 #include "xbt/config.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> smpi_wtime_sleep("smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime",
34 double smpi_cpu_threshold = -1;
35 double smpi_host_speed;
37 SharedMallocType smpi_cfg_shared_malloc = SharedMallocType::GLOBAL;
38 double smpi_total_benched_time = 0;
40 extern "C" XBT_PUBLIC void smpi_execute_flops_(double* flops);
41 void smpi_execute_flops_(double *flops)
43 smpi_execute_flops(*flops);
46 extern "C" XBT_PUBLIC void smpi_execute_(double* duration);
47 void smpi_execute_(double *duration)
49 smpi_execute(*duration);
52 void smpi_execute_flops(double flops) {
53 xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
54 XBT_DEBUG("Handle real computation time: %f flops", flops);
55 simgrid::s4u::this_actor::exec_init(flops)
56 ->set_name("computation")
57 ->set_tracing_category(TRACE_internal_smpi_get_category())
60 smpi_switch_data_segment(simgrid::s4u::Actor::self());
63 void smpi_execute(double duration)
65 if (duration >= smpi_cpu_threshold) {
66 XBT_DEBUG("Sleep for %g to handle real computation time", duration);
67 double flops = duration * smpi_host_speed;
68 int rank = simgrid::s4u::this_actor::get_pid();
69 TRACE_smpi_computing_in(rank, flops);
71 smpi_execute_flops(flops);
73 TRACE_smpi_computing_out(rank);
76 XBT_DEBUG("Real computation took %g while option smpi/cpu-threshold is set to %g => ignore it", duration,
81 void smpi_execute_benched(double duration)
84 double speed = sg_host_speed(sg_host_self());
85 smpi_execute_flops(duration*speed);
89 void smpi_bench_begin()
91 if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
92 smpi_switch_data_segment(simgrid::s4u::Actor::self());
95 if (MC_is_active() || MC_record_replay_is_active())
99 if (not simgrid::config::get_value<std::string>("smpi/papi-events").empty()) {
100 int event_set = smpi_process()->papi_event_set();
101 // PAPI_start sets everything to 0! See man(3) PAPI_start
102 if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized()) {
103 if (PAPI_start(event_set) != PAPI_OK) {
104 // TODO This needs some proper handling.
105 XBT_CRITICAL("Could not start PAPI counters.\n");
111 xbt_os_threadtimer_start(smpi_process()->timer());
114 void smpi_bench_end()
116 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 simgrid::config::get_value<std::string>("smpi/papi-events").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 = std::vector<long long>(counter_data.size());
133 if (PAPI_stop(event_set, &event_values[0]) != PAPI_OK) { // Error
134 XBT_CRITICAL("Could not stop PAPI counters.\n");
137 for (unsigned int i = 0; i < counter_data.size(); i++) {
138 counter_data[i].second += event_values[i];
144 if (smpi_process()->sampling()) {
145 XBT_CRITICAL("Cannot do recursive benchmarks.");
146 XBT_CRITICAL("Are you trying to make a call to MPI within a SMPI_SAMPLE_ block?");
147 xbt_backtrace_display_current();
148 xbt_die("Aborting.");
151 // Maybe we need to artificially speed up or slow down our computation based on our statistical analysis.
152 if (simgrid::config::get_value<std::string>("smpi/comp-adjustment-file")[0] != '\0') {
154 smpi_trace_call_location_t* loc = smpi_process()->call_location();
155 std::string key = loc->get_composed_key();
156 std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
157 if (it != location2speedup.end()) {
158 speedup = it->second;
162 // Simulate the benchmarked computation unless disabled via command-line argument
163 if (simgrid::config::get_value<bool>("smpi/simulate-computation")) {
164 smpi_execute(xbt_os_timer_elapsed(timer)/speedup);
168 if (not simgrid::config::get_value<std::string>("smpi/papi-events").empty() && TRACE_smpi_is_enabled()) {
169 container_t container =
170 simgrid::instr::Container::by_name(std::string("rank-") + std::to_string(simgrid::s4u::this_actor::get_pid()));
171 papi_counter_t& counter_data = smpi_process()->papi_counters();
173 for (auto const& pair : counter_data) {
174 simgrid::instr::VariableType* variable = static_cast<simgrid::instr::VariableType*>(container->type_->by_name(pair.first));
175 variable->set_event(SIMIX_get_clock(), pair.second);
180 smpi_total_benched_time += xbt_os_timer_elapsed(timer);
183 /* Private sleep function used by smpi_sleep(), smpi_usleep() and friends */
184 static unsigned int private_sleep(double secs)
188 XBT_DEBUG("Sleep for: %lf secs", secs);
189 int rank = simgrid::s4u::this_actor::get_pid();
190 TRACE_smpi_sleeping_in(rank, secs);
192 simcall_process_sleep(secs);
194 TRACE_smpi_sleeping_out(rank);
200 unsigned int smpi_sleep(unsigned int secs)
202 if (not smpi_process())
204 return private_sleep(static_cast<double>(secs));
207 int smpi_usleep(useconds_t usecs)
209 if (not smpi_process())
210 return usleep(usecs);
211 return static_cast<int>(private_sleep(static_cast<double>(usecs) / 1000000.0));
214 #if _POSIX_TIMERS > 0
215 int smpi_nanosleep(const struct timespec* tp, struct timespec* t)
217 if (not smpi_process())
218 return nanosleep(tp,t);
219 return static_cast<int>(private_sleep(static_cast<double>(tp->tv_sec + tp->tv_nsec / 1000000000.0)));
223 int smpi_gettimeofday(struct timeval* tv, struct timezone* tz)
225 if (not smpi_process())
226 return gettimeofday(tv, tz);
229 double now = SIMIX_get_clock();
231 tv->tv_sec = static_cast<time_t>(now);
233 tv->tv_usec = static_cast<useconds_t>((now - tv->tv_sec) * 1e6);
235 tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
242 #if _POSIX_TIMERS > 0
243 int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp)
245 if (not smpi_process())
246 return clock_gettime(clk_id, tp);
247 //there is only one time in SMPI, so clk_id is ignored.
249 double now = SIMIX_get_clock();
251 tp->tv_sec = static_cast<time_t>(now);
252 tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
259 extern double sg_surf_precision;
260 unsigned long long smpi_rastro_resolution ()
263 double resolution = (1/sg_surf_precision);
265 return static_cast<unsigned long long>(resolution);
268 unsigned long long smpi_rastro_timestamp ()
271 double now = SIMIX_get_clock();
273 unsigned long long sec = static_cast<unsigned long long>(now);
274 unsigned long long pre = (now - sec) * smpi_rastro_resolution();
276 return static_cast<unsigned long long>(sec) * smpi_rastro_resolution() + pre;
279 double smpi_mpi_wtime()
282 if (smpi_process()->initialized() && not smpi_process()->finalized() && not smpi_process()->sampling()) {
284 time = SIMIX_get_clock();
285 // to avoid deadlocks if used as a break condition, such as
286 // while (MPI_Wtime(...) < time_limit) {
289 // because the time will not normally advance when only calls to MPI_Wtime
290 // are made -> deadlock (MPI_Wtime never reaches the time limit)
291 if (smpi_wtime_sleep > 0)
292 simcall_process_sleep(smpi_wtime_sleep);
295 time = SIMIX_get_clock();
300 /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
302 class SampleLocation : public std::string {
304 SampleLocation(bool global, const char* file, int line) : std::string(std::string(file) + ":" + std::to_string(line))
307 this->append(":" + std::to_string(simgrid::s4u::this_actor::get_pid()));
313 double threshold; /* maximal stderr requested (if positive) */
314 double relstderr; /* observed stderr so far */
315 double mean; /* mean of benched times, to be used if the block is disabled */
316 double sum; /* sum of benched times (to compute the mean and stderr) */
317 double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
318 int iters; /* amount of requested iterations */
319 int count; /* amount of iterations done so far */
320 bool benching; /* true: we are benchmarking; false: we have enough data, no bench anymore */
322 bool need_more_benchs() const;
325 bool LocalData::need_more_benchs() const
327 bool res = (count < iters) || (threshold > 0.0 && (count < 2 || // not enough data
328 relstderr > threshold // stderr too high yet
330 XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
331 (res ? "need more data" : "enough benchs"), count, iters, relstderr, threshold, mean);
335 std::unordered_map<SampleLocation, LocalData, std::hash<std::string>> samples;
338 void smpi_sample_1(int global, const char *file, int line, int iters, double threshold)
340 SampleLocation loc(global, file, line);
342 smpi_bench_end(); /* Take time from previous, unrelated computation into account */
343 smpi_process()->set_sampling(1);
345 auto insert = samples.emplace(loc, LocalData{
346 threshold, // threshold
353 true // benching (if we have no data, we need at least one)
355 LocalData& data = insert.first->second;
357 XBT_DEBUG("XXXXX First time ever on benched nest %s.", loc.c_str());
358 xbt_assert(threshold > 0 || iters > 0,
359 "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
361 if (data.iters != iters || data.threshold != threshold) {
362 XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
363 "How did you manage to give two numbers at the same line??",
364 loc.c_str(), data.iters, data.threshold, iters, threshold);
368 // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate
369 // the computation instead
370 data.benching = data.need_more_benchs();
371 XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc.c_str(),
372 (data.benching ? "more benching needed" : "we have enough data, skip computes"));
376 int smpi_sample_2(int global, const char *file, int line)
378 SampleLocation loc(global, file, line);
381 XBT_DEBUG("sample2 %s", loc.c_str());
382 auto sample = samples.find(loc);
383 if (sample == samples.end())
384 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
385 LocalData& data = sample->second;
388 // we need to run a new bench
389 XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f",
390 data.count, data.iters, data.relstderr, data.threshold, data.mean);
393 // Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
394 //ran one bench and need to bail out now that our job is done). Just sleep instead
395 XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f)."
396 " apply the %fs delay instead",
397 data.count, data.iters, data.relstderr, data.threshold, data.mean);
398 smpi_execute(data.mean);
399 smpi_process()->set_sampling(0);
400 res = 0; // prepare to capture future, unrelated computations
406 void smpi_sample_3(int global, const char *file, int line)
408 SampleLocation loc(global, file, line);
410 XBT_DEBUG("sample3 %s", loc.c_str());
411 auto sample = samples.find(loc);
412 if (sample == samples.end())
413 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
414 LocalData& data = sample->second;
416 if (not data.benching)
419 // ok, benchmarking this loop is over
420 xbt_os_threadtimer_stop(smpi_process()->timer());
424 double period = xbt_os_timer_elapsed(smpi_process()->timer());
426 data.sum_pow2 += period * period;
427 double n = static_cast<double>(data.count);
428 data.mean = data.sum / n;
429 data.relstderr = sqrt((data.sum_pow2 / n - data.mean * data.mean) / n) / data.mean;
430 if (data.need_more_benchs()) {
431 data.mean = period; // Still in benching process; We want sample_2 to simulate the exact time of this loop
432 // occurrence before leaving, not the mean over the history
434 XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)",
435 data.count, data.mean, data.relstderr, period);
437 // That's enough for now, prevent sample_2 to run the same code over and over
438 data.benching = false;
441 smpi_trace_call_location_t* smpi_trace_get_call_location()
443 return smpi_process()->call_location();
446 void smpi_trace_set_call_location(const char* file, const int line)
448 smpi_trace_call_location_t* loc = smpi_process()->call_location();
450 loc->previous_filename = loc->filename;
451 loc->previous_linenumber = loc->linenumber;
452 loc->filename = file;
453 loc->linenumber = line;
456 /** Required for Fortran bindings */
457 void smpi_trace_set_call_location_(const char* file, int* line)
459 smpi_trace_set_call_location(file, *line);
462 /** Required for Fortran if -fsecond-underscore is activated */
463 void smpi_trace_set_call_location__(const char* file, int* line)
465 smpi_trace_set_call_location(file, *line);
468 void smpi_bench_destroy()
473 int smpi_getopt_long_only (int argc, char *const *argv, const char *options,
474 const struct option * long_options, int *opt_index)
477 optind = smpi_process()->get_optind();
478 int ret = getopt_long_only (argc, argv, options, long_options, opt_index);
480 smpi_process()->set_optind(optind);
484 int smpi_getopt_long (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 (argc, argv, options, long_options, opt_index);
491 smpi_process()->set_optind(optind);
495 int smpi_getopt (int argc, char *const *argv, const char *options)
498 optind = smpi_process()->get_optind();
499 int ret = getopt (argc, argv, options);
501 smpi_process()->set_optind(optind);