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 void smpi_execute_flops_(double *flops)
40 smpi_execute_flops(*flops);
43 extern "C" XBT_PUBLIC void smpi_execute_(double* duration);
44 void smpi_execute_(double *duration)
46 smpi_execute(*duration);
49 void smpi_execute_flops(double flops) {
50 xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
51 XBT_DEBUG("Handle real computation time: %f flops", flops);
52 simgrid::s4u::this_actor::exec_init(flops)
53 ->set_name("computation")
54 ->set_tracing_category(TRACE_internal_smpi_get_category())
57 smpi_switch_data_segment(simgrid::s4u::Actor::self());
60 void smpi_execute(double duration)
62 if (duration >= smpi_cpu_threshold) {
63 XBT_DEBUG("Sleep for %g to handle real computation time", duration);
64 double flops = duration * smpi_host_speed;
65 int rank = simgrid::s4u::this_actor::get_pid();
66 TRACE_smpi_computing_in(rank, flops);
68 smpi_execute_flops(flops);
70 TRACE_smpi_computing_out(rank);
73 XBT_DEBUG("Real computation took %g while option smpi/cpu-threshold is set to %g => ignore it", duration,
78 void smpi_execute_benched(double duration)
81 double speed = sg_host_speed(sg_host_self());
82 smpi_execute_flops(duration*speed);
86 void smpi_bench_begin()
88 if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
89 smpi_switch_data_segment(simgrid::s4u::Actor::self());
92 if (MC_is_active() || MC_record_replay_is_active())
96 if (not simgrid::config::get_value<std::string>("smpi/papi-events").empty()) {
97 int event_set = smpi_process()->papi_event_set();
98 // PAPI_start sets everything to 0! See man(3) PAPI_start
99 if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized()) {
100 if (PAPI_start(event_set) != PAPI_OK) {
101 // TODO This needs some proper handling.
102 XBT_CRITICAL("Could not start PAPI counters.\n");
108 xbt_os_threadtimer_start(smpi_process()->timer());
111 void smpi_bench_end()
113 if (MC_is_active() || MC_record_replay_is_active())
117 xbt_os_timer_t timer = smpi_process()->timer();
118 xbt_os_threadtimer_stop(timer);
122 * An MPI function has been called and now is the right time to update
123 * our PAPI counters for this process.
125 if (not simgrid::config::get_value<std::string>("smpi/papi-events").empty()) {
126 papi_counter_t& counter_data = smpi_process()->papi_counters();
127 int event_set = smpi_process()->papi_event_set();
128 std::vector<long long> event_values = std::vector<long long>(counter_data.size());
130 if (PAPI_stop(event_set, &event_values[0]) != PAPI_OK) { // Error
131 XBT_CRITICAL("Could not stop PAPI counters.\n");
134 for (unsigned int i = 0; i < counter_data.size(); i++) {
135 counter_data[i].second += event_values[i];
141 if (smpi_process()->sampling()) {
142 XBT_CRITICAL("Cannot do recursive benchmarks.");
143 XBT_CRITICAL("Are you trying to make a call to MPI within a SMPI_SAMPLE_ block?");
144 xbt_backtrace_display_current();
145 xbt_die("Aborting.");
148 // Maybe we need to artificially speed up or slow down our computation based on our statistical analysis.
149 if (simgrid::config::get_value<std::string>("smpi/comp-adjustment-file")[0] != '\0') {
151 smpi_trace_call_location_t* loc = smpi_process()->call_location();
152 std::string key = loc->get_composed_key();
153 std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
154 if (it != location2speedup.end()) {
155 speedup = it->second;
159 // Simulate the benchmarked computation unless disabled via command-line argument
160 if (simgrid::config::get_value<bool>("smpi/simulate-computation")) {
161 smpi_execute(xbt_os_timer_elapsed(timer)/speedup);
165 if (not simgrid::config::get_value<std::string>("smpi/papi-events").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() and smpi_usleep() */
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 simcall_process_sleep(secs);
191 TRACE_smpi_sleeping_out(rank);
197 unsigned int smpi_sleep(unsigned int secs)
199 if (not smpi_process())
201 return private_sleep(static_cast<double>(secs));
204 int smpi_usleep(useconds_t usecs)
206 if (not smpi_process())
207 return usleep(usecs);
208 return static_cast<int>(private_sleep(static_cast<double>(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(static_cast<double>(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);
239 #if _POSIX_TIMERS > 0
240 int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp)
242 if (not smpi_process())
243 return clock_gettime(clk_id, tp);
244 //there is only one time in SMPI, so clk_id is ignored.
246 double now = SIMIX_get_clock();
248 tp->tv_sec = static_cast<time_t>(now);
249 tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
256 extern double sg_surf_precision;
257 unsigned long long smpi_rastro_resolution ()
260 double resolution = (1/sg_surf_precision);
262 return static_cast<unsigned long long>(resolution);
265 unsigned long long smpi_rastro_timestamp ()
268 double now = SIMIX_get_clock();
270 unsigned long long sec = static_cast<unsigned long long>(now);
271 unsigned long long pre = (now - sec) * smpi_rastro_resolution();
273 return static_cast<unsigned long long>(sec) * smpi_rastro_resolution() + pre;
276 /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
278 class SampleLocation : public std::string {
280 SampleLocation(bool global, const char* file, int line) : std::string(std::string(file) + ":" + std::to_string(line))
283 this->append(":" + std::to_string(simgrid::s4u::this_actor::get_pid()));
289 double threshold; /* maximal stderr requested (if positive) */
290 double relstderr; /* observed stderr so far */
291 double mean; /* mean of benched times, to be used if the block is disabled */
292 double sum; /* sum of benched times (to compute the mean and stderr) */
293 double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
294 int iters; /* amount of requested iterations */
295 int count; /* amount of iterations done so far */
296 bool benching; /* true: we are benchmarking; false: we have enough data, no bench anymore */
298 bool need_more_benchs() const;
301 bool LocalData::need_more_benchs() const
303 bool res = (count < iters) || (threshold > 0.0 && (count < 2 || // not enough data
304 relstderr > threshold // stderr too high yet
306 XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
307 (res ? "need more data" : "enough benchs"), count, iters, relstderr, threshold, mean);
311 std::unordered_map<SampleLocation, LocalData, std::hash<std::string>> samples;
314 void smpi_sample_1(int global, const char *file, int line, int iters, double threshold)
316 SampleLocation loc(global, file, line);
318 smpi_bench_end(); /* Take time from previous, unrelated computation into account */
319 smpi_process()->set_sampling(1);
321 auto insert = samples.emplace(loc, LocalData{
322 threshold, // threshold
329 true // benching (if we have no data, we need at least one)
331 LocalData& data = insert.first->second;
333 XBT_DEBUG("XXXXX First time ever on benched nest %s.", loc.c_str());
334 xbt_assert(threshold > 0 || iters > 0,
335 "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
337 if (data.iters != iters || data.threshold != threshold) {
338 XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
339 "How did you manage to give two numbers at the same line??",
340 loc.c_str(), data.iters, data.threshold, iters, threshold);
344 // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate
345 // the computation instead
346 data.benching = data.need_more_benchs();
347 XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc.c_str(),
348 (data.benching ? "more benching needed" : "we have enough data, skip computes"));
352 int smpi_sample_2(int global, const char *file, int line)
354 SampleLocation loc(global, file, line);
357 XBT_DEBUG("sample2 %s", loc.c_str());
358 auto sample = samples.find(loc);
359 if (sample == samples.end())
360 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
361 LocalData& data = sample->second;
364 // we need to run a new bench
365 XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f",
366 data.count, data.iters, data.relstderr, data.threshold, data.mean);
369 // Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
370 //ran one bench and need to bail out now that our job is done). Just sleep instead
371 XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f)."
372 " apply the %fs delay instead",
373 data.count, data.iters, data.relstderr, data.threshold, data.mean);
374 smpi_execute(data.mean);
375 smpi_process()->set_sampling(0);
376 res = 0; // prepare to capture future, unrelated computations
382 void smpi_sample_3(int global, const char *file, int line)
384 SampleLocation loc(global, file, line);
386 XBT_DEBUG("sample3 %s", loc.c_str());
387 auto sample = samples.find(loc);
388 if (sample == samples.end())
389 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
390 LocalData& data = sample->second;
392 if (not data.benching)
395 // ok, benchmarking this loop is over
396 xbt_os_threadtimer_stop(smpi_process()->timer());
400 double period = xbt_os_timer_elapsed(smpi_process()->timer());
402 data.sum_pow2 += period * period;
403 double n = static_cast<double>(data.count);
404 data.mean = data.sum / n;
405 data.relstderr = sqrt((data.sum_pow2 / n - data.mean * data.mean) / n) / data.mean;
406 if (data.need_more_benchs()) {
407 data.mean = period; // Still in benching process; We want sample_2 to simulate the exact time of this loop
408 // occurrence before leaving, not the mean over the history
410 XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)",
411 data.count, data.mean, data.relstderr, period);
413 // That's enough for now, prevent sample_2 to run the same code over and over
414 data.benching = false;
417 smpi_trace_call_location_t* smpi_trace_get_call_location()
419 return smpi_process()->call_location();
422 void smpi_trace_set_call_location(const char* file, const int line)
424 smpi_trace_call_location_t* loc = smpi_process()->call_location();
426 loc->previous_filename = loc->filename;
427 loc->previous_linenumber = loc->linenumber;
428 loc->filename = file;
429 loc->linenumber = line;
432 /** Required for Fortran bindings */
433 void smpi_trace_set_call_location_(const char* file, int* line)
435 smpi_trace_set_call_location(file, *line);
438 /** Required for Fortran if -fsecond-underscore is activated */
439 void smpi_trace_set_call_location__(const char* file, int* line)
441 smpi_trace_set_call_location(file, *line);
444 void smpi_bench_destroy()
449 int smpi_getopt_long_only (int argc, char *const *argv, const char *options,
450 const struct option * long_options, int *opt_index)
453 optind = smpi_process()->get_optind();
454 int ret = getopt_long_only (argc, argv, options, long_options, opt_index);
456 smpi_process()->set_optind(optind);
460 int smpi_getopt_long (int argc, char *const *argv, const char *options,
461 const struct option * long_options, int *opt_index)
464 optind = smpi_process()->get_optind();
465 int ret = getopt_long (argc, argv, options, long_options, opt_index);
467 smpi_process()->set_optind(optind);
471 int smpi_getopt (int argc, char *const *argv, const char *options)
474 optind = smpi_process()->get_optind();
475 int ret = getopt (argc, argv, options);
477 smpi_process()->set_optind(optind);