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 double smpi_cpu_threshold = -1;
32 double smpi_host_speed;
34 SharedMallocType smpi_cfg_shared_malloc = SharedMallocType::GLOBAL;
35 double smpi_total_benched_time = 0;
37 extern "C" XBT_PUBLIC void smpi_execute_flops_(double* flops);
38 extern "C" XBT_PUBLIC simgrid::config::Flag<double> smpi_wtime_sleep;
40 void smpi_execute_flops_(double *flops)
42 smpi_execute_flops(*flops);
45 extern "C" XBT_PUBLIC void smpi_execute_(double* duration);
46 void smpi_execute_(double *duration)
48 smpi_execute(*duration);
51 void smpi_execute_flops(double flops) {
52 xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
53 XBT_DEBUG("Handle real computation time: %f flops", flops);
54 simgrid::s4u::this_actor::exec_init(flops)
55 ->set_name("computation")
56 ->set_tracing_category(TRACE_internal_smpi_get_category())
59 smpi_switch_data_segment(simgrid::s4u::Actor::self());
62 void smpi_execute(double duration)
64 if (duration >= smpi_cpu_threshold) {
65 XBT_DEBUG("Sleep for %g to handle real computation time", duration);
66 double flops = duration * smpi_host_speed;
67 int rank = simgrid::s4u::this_actor::get_pid();
68 TRACE_smpi_computing_in(rank, flops);
70 smpi_execute_flops(flops);
72 TRACE_smpi_computing_out(rank);
75 XBT_DEBUG("Real computation took %g while option smpi/cpu-threshold is set to %g => ignore it", duration,
80 void smpi_execute_benched(double duration)
83 double speed = sg_host_speed(sg_host_self());
84 smpi_execute_flops(duration*speed);
88 void smpi_bench_begin()
90 if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
91 smpi_switch_data_segment(simgrid::s4u::Actor::self());
94 if (MC_is_active() || MC_record_replay_is_active())
98 if (not simgrid::config::get_value<std::string>("smpi/papi-events").empty()) {
99 int event_set = smpi_process()->papi_event_set();
100 // PAPI_start sets everything to 0! See man(3) PAPI_start
101 if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized()) {
102 if (PAPI_start(event_set) != PAPI_OK) {
103 // TODO This needs some proper handling.
104 XBT_CRITICAL("Could not start PAPI counters.\n");
110 xbt_os_threadtimer_start(smpi_process()->timer());
113 void smpi_bench_end()
115 if (MC_is_active() || MC_record_replay_is_active())
119 xbt_os_timer_t timer = smpi_process()->timer();
120 xbt_os_threadtimer_stop(timer);
124 * An MPI function has been called and now is the right time to update
125 * our PAPI counters for this process.
127 if (not simgrid::config::get_value<std::string>("smpi/papi-events").empty()) {
128 papi_counter_t& counter_data = smpi_process()->papi_counters();
129 int event_set = smpi_process()->papi_event_set();
130 std::vector<long long> event_values = std::vector<long long>(counter_data.size());
132 if (PAPI_stop(event_set, &event_values[0]) != PAPI_OK) { // Error
133 XBT_CRITICAL("Could not stop PAPI counters.\n");
136 for (unsigned int i = 0; i < counter_data.size(); i++) {
137 counter_data[i].second += event_values[i];
143 if (smpi_process()->sampling()) {
144 XBT_CRITICAL("Cannot do recursive benchmarks.");
145 XBT_CRITICAL("Are you trying to make a call to MPI within a SMPI_SAMPLE_ block?");
146 xbt_backtrace_display_current();
147 xbt_die("Aborting.");
150 // Maybe we need to artificially speed up or slow down our computation based on our statistical analysis.
151 if (simgrid::config::get_value<std::string>("smpi/comp-adjustment-file")[0] != '\0') {
153 smpi_trace_call_location_t* loc = smpi_process()->call_location();
154 std::string key = loc->get_composed_key();
155 std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
156 if (it != location2speedup.end()) {
157 speedup = it->second;
161 // Simulate the benchmarked computation unless disabled via command-line argument
162 if (simgrid::config::get_value<bool>("smpi/simulate-computation")) {
163 smpi_execute(xbt_os_timer_elapsed(timer)/speedup);
167 if (not simgrid::config::get_value<std::string>("smpi/papi-events").empty() && TRACE_smpi_is_enabled()) {
168 container_t container =
169 simgrid::instr::Container::by_name(std::string("rank-") + std::to_string(simgrid::s4u::this_actor::get_pid()));
170 papi_counter_t& counter_data = smpi_process()->papi_counters();
172 for (auto const& pair : counter_data) {
173 simgrid::instr::VariableType* variable = static_cast<simgrid::instr::VariableType*>(container->type_->by_name(pair.first));
174 variable->set_event(SIMIX_get_clock(), pair.second);
179 smpi_total_benched_time += xbt_os_timer_elapsed(timer);
182 /* Private sleep function used by smpi_sleep() and smpi_usleep() */
183 static unsigned int private_sleep(double secs)
187 XBT_DEBUG("Sleep for: %lf secs", secs);
188 int rank = simgrid::s4u::this_actor::get_pid();
189 TRACE_smpi_sleeping_in(rank, secs);
191 simcall_process_sleep(secs);
193 TRACE_smpi_sleeping_out(rank);
199 unsigned int smpi_sleep(unsigned int secs)
201 if (not smpi_process())
203 return private_sleep(static_cast<double>(secs));
206 int smpi_usleep(useconds_t usecs)
208 if (not smpi_process())
209 return usleep(usecs);
210 return static_cast<int>(private_sleep(static_cast<double>(usecs) / 1000000.0));
213 #if _POSIX_TIMERS > 0
214 int smpi_nanosleep(const struct timespec* tp, struct timespec* t)
216 if (not smpi_process())
217 return nanosleep(tp,t);
218 return static_cast<int>(private_sleep(static_cast<double>(tp->tv_sec + tp->tv_nsec / 1000000000.0)));
222 int smpi_gettimeofday(struct timeval* tv, struct timezone* tz)
224 if (not smpi_process())
225 return gettimeofday(tv, tz);
228 double now = SIMIX_get_clock();
230 tv->tv_sec = static_cast<time_t>(now);
232 tv->tv_usec = static_cast<useconds_t>((now - tv->tv_sec) * 1e6);
234 tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
237 if(smpi_wtime_sleep > 0)
238 simcall_process_sleep(smpi_wtime_sleep);
243 #if _POSIX_TIMERS > 0
244 int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp)
246 if (not smpi_process())
247 return clock_gettime(clk_id, tp);
248 //there is only one time in SMPI, so clk_id is ignored.
250 double now = SIMIX_get_clock();
252 tp->tv_sec = static_cast<time_t>(now);
253 tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
255 if(smpi_wtime_sleep > 0)
256 simcall_process_sleep(smpi_wtime_sleep);
262 extern double sg_surf_precision;
263 unsigned long long smpi_rastro_resolution ()
266 double resolution = (1/sg_surf_precision);
268 return static_cast<unsigned long long>(resolution);
271 unsigned long long smpi_rastro_timestamp ()
274 double now = SIMIX_get_clock();
276 unsigned long long sec = static_cast<unsigned long long>(now);
277 unsigned long long pre = (now - sec) * smpi_rastro_resolution();
279 return static_cast<unsigned long long>(sec) * smpi_rastro_resolution() + pre;
282 /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
284 class SampleLocation : public std::string {
286 SampleLocation(bool global, const char* file, int line) : std::string(std::string(file) + ":" + std::to_string(line))
289 this->append(":" + std::to_string(simgrid::s4u::this_actor::get_pid()));
295 double threshold; /* maximal stderr requested (if positive) */
296 double relstderr; /* observed stderr so far */
297 double mean; /* mean of benched times, to be used if the block is disabled */
298 double sum; /* sum of benched times (to compute the mean and stderr) */
299 double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
300 int iters; /* amount of requested iterations */
301 int count; /* amount of iterations done so far */
302 bool benching; /* true: we are benchmarking; false: we have enough data, no bench anymore */
304 bool need_more_benchs() const;
307 bool LocalData::need_more_benchs() const
309 bool res = (count < iters) || (threshold > 0.0 && (count < 2 || // not enough data
310 relstderr > threshold // stderr too high yet
312 XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
313 (res ? "need more data" : "enough benchs"), count, iters, relstderr, threshold, mean);
317 std::unordered_map<SampleLocation, LocalData, std::hash<std::string>> samples;
320 void smpi_sample_1(int global, const char *file, int line, int iters, double threshold)
322 SampleLocation loc(global, file, line);
324 smpi_bench_end(); /* Take time from previous, unrelated computation into account */
325 smpi_process()->set_sampling(1);
327 auto insert = samples.emplace(loc, LocalData{
328 threshold, // threshold
335 true // benching (if we have no data, we need at least one)
337 LocalData& data = insert.first->second;
339 XBT_DEBUG("XXXXX First time ever on benched nest %s.", loc.c_str());
340 xbt_assert(threshold > 0 || iters > 0,
341 "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
343 if (data.iters != iters || data.threshold != threshold) {
344 XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
345 "How did you manage to give two numbers at the same line??",
346 loc.c_str(), data.iters, data.threshold, iters, threshold);
350 // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate
351 // the computation instead
352 data.benching = data.need_more_benchs();
353 XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc.c_str(),
354 (data.benching ? "more benching needed" : "we have enough data, skip computes"));
358 int smpi_sample_2(int global, const char *file, int line)
360 SampleLocation loc(global, file, line);
363 XBT_DEBUG("sample2 %s", loc.c_str());
364 auto sample = samples.find(loc);
365 if (sample == samples.end())
366 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
367 LocalData& data = sample->second;
370 // we need to run a new bench
371 XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f",
372 data.count, data.iters, data.relstderr, data.threshold, data.mean);
375 // Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
376 //ran one bench and need to bail out now that our job is done). Just sleep instead
377 XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f)."
378 " apply the %fs delay instead",
379 data.count, data.iters, data.relstderr, data.threshold, data.mean);
380 smpi_execute(data.mean);
381 smpi_process()->set_sampling(0);
382 res = 0; // prepare to capture future, unrelated computations
388 void smpi_sample_3(int global, const char *file, int line)
390 SampleLocation loc(global, file, line);
392 XBT_DEBUG("sample3 %s", loc.c_str());
393 auto sample = samples.find(loc);
394 if (sample == samples.end())
395 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
396 LocalData& data = sample->second;
398 if (not data.benching)
401 // ok, benchmarking this loop is over
402 xbt_os_threadtimer_stop(smpi_process()->timer());
406 double period = xbt_os_timer_elapsed(smpi_process()->timer());
408 data.sum_pow2 += period * period;
409 double n = static_cast<double>(data.count);
410 data.mean = data.sum / n;
411 data.relstderr = sqrt((data.sum_pow2 / n - data.mean * data.mean) / n) / data.mean;
412 if (data.need_more_benchs()) {
413 data.mean = period; // Still in benching process; We want sample_2 to simulate the exact time of this loop
414 // occurrence before leaving, not the mean over the history
416 XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)",
417 data.count, data.mean, data.relstderr, period);
419 // That's enough for now, prevent sample_2 to run the same code over and over
420 data.benching = false;
423 smpi_trace_call_location_t* smpi_trace_get_call_location()
425 return smpi_process()->call_location();
428 void smpi_trace_set_call_location(const char* file, const int line)
430 smpi_trace_call_location_t* loc = smpi_process()->call_location();
432 loc->previous_filename = loc->filename;
433 loc->previous_linenumber = loc->linenumber;
434 loc->filename = file;
435 loc->linenumber = line;
438 /** Required for Fortran bindings */
439 void smpi_trace_set_call_location_(const char* file, 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, 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);