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. */
7 #include "simgrid/host.h"
8 #include "simgrid/modelchecker.h"
9 #include "smpi_comm.hpp"
10 #include "smpi_process.hpp"
11 #include "src/internal_config.h"
12 #include "src/mc/mc_replay.hpp"
13 #include "src/simix/ActorImpl.hpp"
14 #include "xbt/config.hpp"
17 #include <unordered_map>
28 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)");
30 double smpi_cpu_threshold = -1;
31 double smpi_host_speed;
33 shared_malloc_type smpi_cfg_shared_malloc = shmalloc_global;
34 double smpi_total_benched_time = 0;
36 extern "C" XBT_PUBLIC void smpi_execute_flops_(double* flops);
37 void smpi_execute_flops_(double *flops)
39 smpi_execute_flops(*flops);
42 extern "C" XBT_PUBLIC void smpi_execute_(double* duration);
43 void smpi_execute_(double *duration)
45 smpi_execute(*duration);
48 void smpi_execute_flops(double flops) {
49 xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
50 XBT_DEBUG("Handle real computation time: %f flops", flops);
51 smx_activity_t action = simcall_execution_start("computation", flops, 1, 0, smpi_process()->process()->get_host());
52 simcall_set_category (action, TRACE_internal_smpi_get_category());
53 simcall_execution_wait(action);
54 smpi_switch_data_segment(simgrid::s4u::Actor::self());
57 void smpi_execute(double duration)
59 if (duration >= smpi_cpu_threshold) {
60 XBT_DEBUG("Sleep for %g to handle real computation time", duration);
61 double flops = duration * smpi_host_speed;
62 int rank = simgrid::s4u::this_actor::get_pid();
63 TRACE_smpi_computing_in(rank, flops);
65 smpi_execute_flops(flops);
67 TRACE_smpi_computing_out(rank);
70 XBT_DEBUG("Real computation took %g while option smpi/cpu-threshold is set to %g => ignore it", duration,
75 void smpi_execute_benched(double duration)
78 double speed = sg_host_speed(sg_host_self());
79 smpi_execute_flops(duration*speed);
83 void smpi_bench_begin()
85 if (smpi_privatize_global_variables == SmpiPrivStrategies::Mmap) {
86 smpi_switch_data_segment(simgrid::s4u::Actor::self());
89 if (MC_is_active() || MC_record_replay_is_active())
93 if (not xbt_cfg_get_string("smpi/papi-events").empty()) {
94 int event_set = smpi_process()->papi_event_set();
95 // PAPI_start sets everything to 0! See man(3) PAPI_start
96 if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized()) {
97 if (PAPI_start(event_set) != PAPI_OK) {
98 // TODO This needs some proper handling.
99 XBT_CRITICAL("Could not start PAPI counters.\n");
105 xbt_os_threadtimer_start(smpi_process()->timer());
108 void smpi_bench_end()
110 if (MC_is_active() || MC_record_replay_is_active())
114 xbt_os_timer_t timer = smpi_process()->timer();
115 xbt_os_threadtimer_stop(timer);
119 * An MPI function has been called and now is the right time to update
120 * our PAPI counters for this process.
122 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
123 papi_counter_t& counter_data = smpi_process()->papi_counters();
124 int event_set = smpi_process()->papi_event_set();
125 std::vector<long long> event_values = std::vector<long long>(counter_data.size());
127 if (PAPI_stop(event_set, &event_values[0]) != PAPI_OK) { // Error
128 XBT_CRITICAL("Could not stop PAPI counters.\n");
131 for (unsigned int i = 0; i < counter_data.size(); i++) {
132 counter_data[i].second += event_values[i];
138 if (smpi_process()->sampling()) {
139 XBT_CRITICAL("Cannot do recursive benchmarks.");
140 XBT_CRITICAL("Are you trying to make a call to MPI within a SMPI_SAMPLE_ block?");
141 xbt_backtrace_display_current();
142 xbt_die("Aborting.");
145 if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') { // Maybe we need to artificially speed up or slow
146 // down our computation based on our statistical analysis.
148 smpi_trace_call_location_t* loc = smpi_process()->call_location();
149 std::string key = loc->get_composed_key();
150 std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
151 if (it != location2speedup.end()) {
152 speedup = it->second;
156 // Simulate the benchmarked computation unless disabled via command-line argument
157 if (xbt_cfg_get_boolean("smpi/simulate-computation")) {
158 smpi_execute(xbt_os_timer_elapsed(timer)/speedup);
162 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0' && TRACE_smpi_is_enabled()) {
163 container_t container =
164 new simgrid::instr::Container(std::string("rank-") + std::to_string(simgrid::s4u::this_actor::get_pid));
165 papi_counter_t& counter_data = smpi_process()->papi_counters();
167 for (auto const& pair : counter_data) {
168 new simgrid::instr::SetVariableEvent(
169 surf_get_clock(), container, PJ_type_get(/* countername */ pair.first.c_str(), container->type), pair.second);
174 smpi_total_benched_time += xbt_os_timer_elapsed(timer);
177 /* Private sleep function used by smpi_sleep() and smpi_usleep() */
178 static unsigned int private_sleep(double secs)
182 XBT_DEBUG("Sleep for: %lf secs", secs);
183 int rank = MPI_COMM_WORLD->rank();
184 TRACE_smpi_sleeping_in(rank, secs);
186 simcall_process_sleep(secs);
188 TRACE_smpi_sleeping_out(rank);
194 unsigned int smpi_sleep(unsigned int secs)
196 if (not smpi_process())
198 return private_sleep(static_cast<double>(secs));
201 int smpi_usleep(useconds_t usecs)
203 if (not smpi_process())
204 return usleep(usecs);
205 return static_cast<int>(private_sleep(static_cast<double>(usecs) / 1000000.0));
208 #if _POSIX_TIMERS > 0
209 int smpi_nanosleep(const struct timespec* tp, struct timespec* t)
211 if (not smpi_process())
212 return nanosleep(tp,t);
213 return static_cast<int>(private_sleep(static_cast<double>(tp->tv_sec + tp->tv_nsec / 1000000000.0)));
217 int smpi_gettimeofday(struct timeval* tv, struct timezone* tz)
219 if (not smpi_process())
220 return gettimeofday(tv, tz);
223 double now = SIMIX_get_clock();
225 tv->tv_sec = static_cast<time_t>(now);
227 tv->tv_usec = static_cast<useconds_t>((now - tv->tv_sec) * 1e6);
229 tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
236 #if _POSIX_TIMERS > 0
237 int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp)
239 if (not smpi_process())
240 return clock_gettime(clk_id, tp);
241 //there is only one time in SMPI, so clk_id is ignored.
243 double now = SIMIX_get_clock();
245 tp->tv_sec = static_cast<time_t>(now);
246 tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
253 extern double sg_surf_precision;
254 unsigned long long smpi_rastro_resolution ()
257 double resolution = (1/sg_surf_precision);
259 return static_cast<unsigned long long>(resolution);
262 unsigned long long smpi_rastro_timestamp ()
265 double now = SIMIX_get_clock();
267 unsigned long long sec = static_cast<unsigned long long>(now);
268 unsigned long long pre = (now - sec) * smpi_rastro_resolution();
270 return static_cast<unsigned long long>(sec) * smpi_rastro_resolution() + pre;
273 /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
275 class SampleLocation : public std::string {
277 SampleLocation(bool global, const char* file, int line) : std::string(std::string(file) + ":" + std::to_string(line))
280 this->append(":" + std::to_string(simgrid::s4u::this_actor::get_pid()));
286 double threshold; /* maximal stderr requested (if positive) */
287 double relstderr; /* observed stderr so far */
288 double mean; /* mean of benched times, to be used if the block is disabled */
289 double sum; /* sum of benched times (to compute the mean and stderr) */
290 double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
291 int iters; /* amount of requested iterations */
292 int count; /* amount of iterations done so far */
293 bool benching; /* true: we are benchmarking; false: we have enough data, no bench anymore */
295 bool need_more_benchs() const;
299 std::unordered_map<SampleLocation, LocalData, std::hash<std::string>> samples;
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 void smpi_sample_1(int global, const char *file, int line, int iters, double threshold)
313 SampleLocation loc(global, file, line);
315 smpi_bench_end(); /* Take time from previous, unrelated computation into account */
316 smpi_process()->set_sampling(1);
318 auto insert = samples.emplace(loc, LocalData{
319 threshold, // threshold
326 true // benching (if we have no data, we need at least one)
328 LocalData& data = insert.first->second;
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 if (data.iters != iters || data.threshold != threshold) {
335 XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
336 "How did you manage to give two numbers at the same line??",
337 loc.c_str(), data.iters, data.threshold, iters, threshold);
341 // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate
342 // the computation instead
343 data.benching = data.need_more_benchs();
344 XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc.c_str(),
345 (data.benching ? "more benching needed" : "we have enough data, skip computes"));
349 int smpi_sample_2(int global, const char *file, int line)
351 SampleLocation loc(global, file, line);
354 XBT_DEBUG("sample2 %s", loc.c_str());
355 auto sample = samples.find(loc);
356 if (sample == samples.end())
357 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
358 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",
363 data.count, data.iters, data.relstderr, data.threshold, data.mean);
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 XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f)."
369 " apply the %fs delay instead",
370 data.count, data.iters, data.relstderr, data.threshold, data.mean);
371 smpi_execute(data.mean);
372 smpi_process()->set_sampling(0);
373 res = 0; // prepare to capture future, unrelated computations
379 void smpi_sample_3(int global, const char *file, int line)
381 SampleLocation loc(global, file, line);
383 XBT_DEBUG("sample3 %s", loc.c_str());
384 auto sample = samples.find(loc);
385 if (sample == samples.end())
386 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
387 LocalData& data = sample->second;
389 if (not data.benching)
392 // ok, benchmarking this loop is over
393 xbt_os_threadtimer_stop(smpi_process()->timer());
397 double period = xbt_os_timer_elapsed(smpi_process()->timer());
399 data.sum_pow2 += period * period;
400 double n = static_cast<double>(data.count);
401 data.mean = data.sum / n;
402 data.relstderr = sqrt((data.sum_pow2 / n - data.mean * data.mean) / n) / data.mean;
403 if (data.need_more_benchs()) {
404 data.mean = period; // Still in benching process; We want sample_2 to simulate the exact time of this loop
405 // occurrence before leaving, not the mean over the history
407 XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)",
408 data.count, data.mean, data.relstderr, period);
410 // That's enough for now, prevent sample_2 to run the same code over and over
411 data.benching = false;
414 smpi_trace_call_location_t* smpi_trace_get_call_location()
416 return smpi_process()->call_location();
419 void smpi_trace_set_call_location(const char* file, const int line)
421 smpi_trace_call_location_t* loc = smpi_process()->call_location();
423 loc->previous_filename = loc->filename;
424 loc->previous_linenumber = loc->linenumber;
425 loc->filename = file;
426 loc->linenumber = line;
429 /** Required for Fortran bindings */
430 void smpi_trace_set_call_location_(const char* file, int* line)
432 smpi_trace_set_call_location(file, *line);
435 /** Required for Fortran if -fsecond-underscore is activated */
436 void smpi_trace_set_call_location__(const char* file, int* line)
438 smpi_trace_set_call_location(file, *line);
441 void smpi_bench_destroy()
446 int smpi_getopt_long (int argc, char *const *argv, const char *options,
447 const struct option * long_options, int *opt_index)
450 optind = smpi_process()->get_optind();
451 int ret = getopt_long (argc, argv, options, long_options, opt_index);
453 smpi_process()->set_optind(optind);
457 int smpi_getopt (int argc, char *const *argv, const char *options)
460 optind = smpi_process()->get_optind();
461 int ret = getopt (argc, argv, options);
463 smpi_process()->set_optind(optind);