1 /* Copyright (c) 2007, 2009-2018. 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 <unordered_map>
25 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)");
27 double smpi_cpu_threshold = -1;
28 double smpi_host_speed;
30 shared_malloc_type smpi_cfg_shared_malloc = shmalloc_global;
31 double smpi_total_benched_time = 0;
33 extern "C" XBT_PUBLIC void smpi_execute_flops_(double* flops);
34 void smpi_execute_flops_(double *flops)
36 smpi_execute_flops(*flops);
39 extern "C" XBT_PUBLIC void smpi_execute_(double* duration);
40 void smpi_execute_(double *duration)
42 smpi_execute(*duration);
45 void smpi_execute_flops(double flops) {
46 xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
47 XBT_DEBUG("Handle real computation time: %f flops", flops);
48 smx_activity_t action = simcall_execution_start("computation", flops, 1, 0, smpi_process()->process()->getHost());
49 simcall_set_category (action, TRACE_internal_smpi_get_category());
50 simcall_execution_wait(action);
51 smpi_switch_data_segment(simgrid::s4u::Actor::self());
54 void smpi_execute(double duration)
56 if (duration >= smpi_cpu_threshold) {
57 XBT_DEBUG("Sleep for %g to handle real computation time", duration);
58 double flops = duration * smpi_host_speed;
59 int rank = simgrid::s4u::this_actor::getPid();
60 TRACE_smpi_computing_in(rank, flops);
62 smpi_execute_flops(flops);
64 TRACE_smpi_computing_out(rank);
67 XBT_DEBUG("Real computation took %g while option smpi/cpu-threshold is set to %g => ignore it", duration,
72 void smpi_execute_benched(double duration)
75 double speed = sg_host_speed(sg_host_self());
76 smpi_execute_flops(duration*speed);
80 void smpi_bench_begin()
82 if (smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) {
83 smpi_switch_data_segment(simgrid::s4u::Actor::self());
86 if (MC_is_active() || MC_record_replay_is_active())
90 if (not xbt_cfg_get_string("smpi/papi-events").empty()) {
91 int event_set = smpi_process()->papi_event_set();
92 // PAPI_start sets everything to 0! See man(3) PAPI_start
93 if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized()) {
94 if (PAPI_start(event_set) != PAPI_OK) {
95 // TODO This needs some proper handling.
96 XBT_CRITICAL("Could not start PAPI counters.\n");
102 xbt_os_threadtimer_start(smpi_process()->timer());
105 void smpi_bench_end()
107 if (MC_is_active() || MC_record_replay_is_active())
111 xbt_os_timer_t timer = smpi_process()->timer();
112 xbt_os_threadtimer_stop(timer);
116 * An MPI function has been called and now is the right time to update
117 * our PAPI counters for this process.
119 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
120 papi_counter_t& counter_data = smpi_process()->papi_counters();
121 int event_set = smpi_process()->papi_event_set();
122 std::vector<long long> event_values = std::vector<long long>(counter_data.size());
124 if (PAPI_stop(event_set, &event_values[0]) != PAPI_OK) { // Error
125 XBT_CRITICAL("Could not stop PAPI counters.\n");
128 for (unsigned int i = 0; i < counter_data.size(); i++) {
129 counter_data[i].second += event_values[i];
135 if (smpi_process()->sampling()) {
136 XBT_CRITICAL("Cannot do recursive benchmarks.");
137 XBT_CRITICAL("Are you trying to make a call to MPI within a SMPI_SAMPLE_ block?");
138 xbt_backtrace_display_current();
139 xbt_die("Aborting.");
142 if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') { // Maybe we need to artificially speed up or slow
143 // down our computation based on our statistical analysis.
145 smpi_trace_call_location_t* loc = smpi_process()->call_location();
146 std::string key = loc->get_composed_key();
147 std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
148 if (it != location2speedup.end()) {
149 speedup = it->second;
153 // Simulate the benchmarked computation unless disabled via command-line argument
154 if (xbt_cfg_get_boolean("smpi/simulate-computation")) {
155 smpi_execute(xbt_os_timer_elapsed(timer)/speedup);
159 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0' && TRACE_smpi_is_enabled()) {
160 container_t container =
161 new simgrid::instr::Container(std::string("rank-") + std::to_string(simgrid::s4u::this_actor::getPid));
162 papi_counter_t& counter_data = smpi_process()->papi_counters();
164 for (auto const& pair : counter_data) {
165 new simgrid::instr::SetVariableEvent(
166 surf_get_clock(), container, PJ_type_get(/* countername */ pair.first.c_str(), container->type), pair.second);
171 smpi_total_benched_time += xbt_os_timer_elapsed(timer);
174 /* Private sleep function used by smpi_sleep() and smpi_usleep() */
175 static unsigned int private_sleep(double secs)
179 XBT_DEBUG("Sleep for: %lf secs", secs);
180 int rank = MPI_COMM_WORLD->rank();
181 TRACE_smpi_sleeping_in(rank, secs);
183 simcall_process_sleep(secs);
185 TRACE_smpi_sleeping_out(rank);
191 unsigned int smpi_sleep(unsigned int secs)
193 return private_sleep(static_cast<double>(secs));
196 int smpi_usleep(useconds_t usecs)
198 return static_cast<int>(private_sleep(static_cast<double>(usecs) / 1000000.0));
201 #if _POSIX_TIMERS > 0
202 int smpi_nanosleep(const struct timespec* tp, struct timespec* /*t*/)
204 return static_cast<int>(private_sleep(static_cast<double>(tp->tv_sec + tp->tv_nsec / 1000000000.0)));
208 int smpi_gettimeofday(struct timeval* tv, void* /*tz*/)
211 double now = SIMIX_get_clock();
213 tv->tv_sec = static_cast<time_t>(now);
215 tv->tv_usec = static_cast<useconds_t>((now - tv->tv_sec) * 1e6);
217 tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
224 #if _POSIX_TIMERS > 0
225 int smpi_clock_gettime(clockid_t /*clk_id*/, struct timespec* tp)
227 //there is only one time in SMPI, so clk_id is ignored.
229 double now = SIMIX_get_clock();
231 tp->tv_sec = static_cast<time_t>(now);
232 tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
239 extern double sg_surf_precision;
240 unsigned long long smpi_rastro_resolution ()
243 double resolution = (1/sg_surf_precision);
245 return static_cast<unsigned long long>(resolution);
248 unsigned long long smpi_rastro_timestamp ()
251 double now = SIMIX_get_clock();
253 unsigned long long sec = static_cast<unsigned long long>(now);
254 unsigned long long pre = (now - sec) * smpi_rastro_resolution();
256 return static_cast<unsigned long long>(sec) * smpi_rastro_resolution() + pre;
259 /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
261 class SampleLocation : public std::string {
263 SampleLocation(bool global, const char* file, int line) : std::string(std::string(file) + ":" + std::to_string(line))
266 this->append(":" + std::to_string(simgrid::s4u::this_actor::getPid()));
272 double threshold; /* maximal stderr requested (if positive) */
273 double relstderr; /* observed stderr so far */
274 double mean; /* mean of benched times, to be used if the block is disabled */
275 double sum; /* sum of benched times (to compute the mean and stderr) */
276 double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
277 int iters; /* amount of requested iterations */
278 int count; /* amount of iterations done so far */
279 bool benching; /* true: we are benchmarking; false: we have enough data, no bench anymore */
281 bool need_more_benchs() const;
285 std::unordered_map<SampleLocation, LocalData, std::hash<std::string>> samples;
287 bool LocalData::need_more_benchs() const
289 bool res = (count < iters) || (threshold > 0.0 && (count < 2 || // not enough data
290 relstderr > threshold // stderr too high yet
292 XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
293 (res ? "need more data" : "enough benchs"), count, iters, relstderr, threshold, mean);
297 void smpi_sample_1(int global, const char *file, int line, int iters, double threshold)
299 SampleLocation loc(global, file, line);
301 smpi_bench_end(); /* Take time from previous, unrelated computation into account */
302 smpi_process()->set_sampling(1);
304 auto insert = samples.emplace(loc, LocalData{
305 threshold, // threshold
312 true // benching (if we have no data, we need at least one)
314 LocalData& data = insert.first->second;
316 XBT_DEBUG("XXXXX First time ever on benched nest %s.", loc.c_str());
317 xbt_assert(threshold > 0 || iters > 0,
318 "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
320 if (data.iters != iters || data.threshold != threshold) {
321 XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
322 "How did you manage to give two numbers at the same line??",
323 loc.c_str(), data.iters, data.threshold, iters, threshold);
327 // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate
328 // the computation instead
329 data.benching = data.need_more_benchs();
330 XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc.c_str(),
331 (data.benching ? "more benching needed" : "we have enough data, skip computes"));
335 int smpi_sample_2(int global, const char *file, int line)
337 SampleLocation loc(global, file, line);
340 XBT_DEBUG("sample2 %s", loc.c_str());
341 auto sample = samples.find(loc);
342 if (sample == samples.end())
343 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
344 LocalData& data = sample->second;
347 // we need to run a new bench
348 XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f",
349 data.count, data.iters, data.relstderr, data.threshold, data.mean);
352 // Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
353 //ran one bench and need to bail out now that our job is done). Just sleep instead
354 XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f)."
355 " apply the %fs delay instead",
356 data.count, data.iters, data.relstderr, data.threshold, data.mean);
357 smpi_execute(data.mean);
358 smpi_process()->set_sampling(0);
359 res = 0; // prepare to capture future, unrelated computations
365 void smpi_sample_3(int global, const char *file, int line)
367 SampleLocation loc(global, file, line);
369 XBT_DEBUG("sample3 %s", loc.c_str());
370 auto sample = samples.find(loc);
371 if (sample == samples.end())
372 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
373 LocalData& data = sample->second;
375 if (not data.benching)
378 // ok, benchmarking this loop is over
379 xbt_os_threadtimer_stop(smpi_process()->timer());
383 double period = xbt_os_timer_elapsed(smpi_process()->timer());
385 data.sum_pow2 += period * period;
386 double n = static_cast<double>(data.count);
387 data.mean = data.sum / n;
388 data.relstderr = sqrt((data.sum_pow2 / n - data.mean * data.mean) / n) / data.mean;
389 if (data.need_more_benchs()) {
390 data.mean = period; // Still in benching process; We want sample_2 to simulate the exact time of this loop
391 // occurrence before leaving, not the mean over the history
393 XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)",
394 data.count, data.mean, data.relstderr, period);
396 // That's enough for now, prevent sample_2 to run the same code over and over
397 data.benching = false;
400 smpi_trace_call_location_t* smpi_trace_get_call_location()
402 return smpi_process()->call_location();
405 void smpi_trace_set_call_location(const char* file, const int line)
407 smpi_trace_call_location_t* loc = smpi_process()->call_location();
409 loc->previous_filename = loc->filename;
410 loc->previous_linenumber = loc->linenumber;
411 loc->filename = file;
412 loc->linenumber = line;
415 /** Required for Fortran bindings */
416 void smpi_trace_set_call_location_(const char* file, int* line)
418 smpi_trace_set_call_location(file, *line);
421 /** Required for Fortran if -fsecond-underscore is activated */
422 void smpi_trace_set_call_location__(const char* file, int* line)
424 smpi_trace_set_call_location(file, *line);
427 void smpi_bench_destroy()