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 "xbt/config.hpp"
16 #include <unordered_map>
27 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)");
29 double smpi_cpu_threshold = -1;
30 double smpi_host_speed;
32 shared_malloc_type smpi_cfg_shared_malloc = shmalloc_global;
33 double smpi_total_benched_time = 0;
35 extern "C" XBT_PUBLIC void smpi_execute_flops_(double* flops);
36 void smpi_execute_flops_(double *flops)
38 smpi_execute_flops(*flops);
41 extern "C" XBT_PUBLIC void smpi_execute_(double* duration);
42 void smpi_execute_(double *duration)
44 smpi_execute(*duration);
47 void smpi_execute_flops(double flops) {
48 xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
49 XBT_DEBUG("Handle real computation time: %f flops", flops);
50 smx_activity_t action = simcall_execution_start("computation", flops, 1, 0, smpi_process()->process()->getHost());
51 simcall_set_category (action, TRACE_internal_smpi_get_category());
52 simcall_execution_wait(action);
53 smpi_switch_data_segment(simgrid::s4u::Actor::self());
56 void smpi_execute(double duration)
58 if (duration >= smpi_cpu_threshold) {
59 XBT_DEBUG("Sleep for %g to handle real computation time", duration);
60 double flops = duration * smpi_host_speed;
61 int rank = simgrid::s4u::this_actor::getPid();
62 TRACE_smpi_computing_in(rank, flops);
64 smpi_execute_flops(flops);
66 TRACE_smpi_computing_out(rank);
69 XBT_DEBUG("Real computation took %g while option smpi/cpu-threshold is set to %g => ignore it", duration,
74 void smpi_execute_benched(double duration)
77 double speed = sg_host_speed(sg_host_self());
78 smpi_execute_flops(duration*speed);
82 void smpi_bench_begin()
84 if (smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) {
85 smpi_switch_data_segment(simgrid::s4u::Actor::self());
88 if (MC_is_active() || MC_record_replay_is_active())
92 if (not xbt_cfg_get_string("smpi/papi-events").empty()) {
93 int event_set = smpi_process()->papi_event_set();
94 // PAPI_start sets everything to 0! See man(3) PAPI_start
95 if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized()) {
96 if (PAPI_start(event_set) != PAPI_OK) {
97 // TODO This needs some proper handling.
98 XBT_CRITICAL("Could not start PAPI counters.\n");
104 xbt_os_threadtimer_start(smpi_process()->timer());
107 void smpi_bench_end()
109 if (MC_is_active() || MC_record_replay_is_active())
113 xbt_os_timer_t timer = smpi_process()->timer();
114 xbt_os_threadtimer_stop(timer);
118 * An MPI function has been called and now is the right time to update
119 * our PAPI counters for this process.
121 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
122 papi_counter_t& counter_data = smpi_process()->papi_counters();
123 int event_set = smpi_process()->papi_event_set();
124 std::vector<long long> event_values = std::vector<long long>(counter_data.size());
126 if (PAPI_stop(event_set, &event_values[0]) != PAPI_OK) { // Error
127 XBT_CRITICAL("Could not stop PAPI counters.\n");
130 for (unsigned int i = 0; i < counter_data.size(); i++) {
131 counter_data[i].second += event_values[i];
137 if (smpi_process()->sampling()) {
138 XBT_CRITICAL("Cannot do recursive benchmarks.");
139 XBT_CRITICAL("Are you trying to make a call to MPI within a SMPI_SAMPLE_ block?");
140 xbt_backtrace_display_current();
141 xbt_die("Aborting.");
144 if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') { // Maybe we need to artificially speed up or slow
145 // down our computation based on our statistical analysis.
147 smpi_trace_call_location_t* loc = smpi_process()->call_location();
148 std::string key = loc->get_composed_key();
149 std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
150 if (it != location2speedup.end()) {
151 speedup = it->second;
155 // Simulate the benchmarked computation unless disabled via command-line argument
156 if (xbt_cfg_get_boolean("smpi/simulate-computation")) {
157 smpi_execute(xbt_os_timer_elapsed(timer)/speedup);
161 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0' && TRACE_smpi_is_enabled()) {
162 container_t container =
163 new simgrid::instr::Container(std::string("rank-") + std::to_string(simgrid::s4u::this_actor::getPid));
164 papi_counter_t& counter_data = smpi_process()->papi_counters();
166 for (auto const& pair : counter_data) {
167 new simgrid::instr::SetVariableEvent(
168 surf_get_clock(), container, PJ_type_get(/* countername */ pair.first.c_str(), container->type), pair.second);
173 smpi_total_benched_time += xbt_os_timer_elapsed(timer);
176 /* Private sleep function used by smpi_sleep() and smpi_usleep() */
177 static unsigned int private_sleep(double secs)
181 XBT_DEBUG("Sleep for: %lf secs", secs);
182 int rank = MPI_COMM_WORLD->rank();
183 TRACE_smpi_sleeping_in(rank, secs);
185 simcall_process_sleep(secs);
187 TRACE_smpi_sleeping_out(rank);
193 unsigned int smpi_sleep(unsigned int secs)
195 return private_sleep(static_cast<double>(secs));
198 int smpi_usleep(useconds_t usecs)
200 return static_cast<int>(private_sleep(static_cast<double>(usecs) / 1000000.0));
203 #if _POSIX_TIMERS > 0
204 int smpi_nanosleep(const struct timespec* tp, struct timespec* /*t*/)
206 return static_cast<int>(private_sleep(static_cast<double>(tp->tv_sec + tp->tv_nsec / 1000000000.0)));
210 int smpi_gettimeofday(struct timeval* tv, void* /*tz*/)
213 double now = SIMIX_get_clock();
215 tv->tv_sec = static_cast<time_t>(now);
217 tv->tv_usec = static_cast<useconds_t>((now - tv->tv_sec) * 1e6);
219 tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
226 #if _POSIX_TIMERS > 0
227 int smpi_clock_gettime(clockid_t /*clk_id*/, struct timespec* tp)
229 //there is only one time in SMPI, so clk_id is ignored.
231 double now = SIMIX_get_clock();
233 tp->tv_sec = static_cast<time_t>(now);
234 tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
241 extern double sg_surf_precision;
242 unsigned long long smpi_rastro_resolution ()
245 double resolution = (1/sg_surf_precision);
247 return static_cast<unsigned long long>(resolution);
250 unsigned long long smpi_rastro_timestamp ()
253 double now = SIMIX_get_clock();
255 unsigned long long sec = static_cast<unsigned long long>(now);
256 unsigned long long pre = (now - sec) * smpi_rastro_resolution();
258 return static_cast<unsigned long long>(sec) * smpi_rastro_resolution() + pre;
261 /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
263 class SampleLocation : public std::string {
265 SampleLocation(bool global, const char* file, int line) : std::string(std::string(file) + ":" + std::to_string(line))
268 this->append(":" + std::to_string(simgrid::s4u::this_actor::getPid()));
274 double threshold; /* maximal stderr requested (if positive) */
275 double relstderr; /* observed stderr so far */
276 double mean; /* mean of benched times, to be used if the block is disabled */
277 double sum; /* sum of benched times (to compute the mean and stderr) */
278 double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
279 int iters; /* amount of requested iterations */
280 int count; /* amount of iterations done so far */
281 bool benching; /* true: we are benchmarking; false: we have enough data, no bench anymore */
283 bool need_more_benchs() const;
287 std::unordered_map<SampleLocation, LocalData, std::hash<std::string>> samples;
289 bool LocalData::need_more_benchs() const
291 bool res = (count < iters) || (threshold > 0.0 && (count < 2 || // not enough data
292 relstderr > threshold // stderr too high yet
294 XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
295 (res ? "need more data" : "enough benchs"), count, iters, relstderr, threshold, mean);
299 void smpi_sample_1(int global, const char *file, int line, int iters, double threshold)
301 SampleLocation loc(global, file, line);
303 smpi_bench_end(); /* Take time from previous, unrelated computation into account */
304 smpi_process()->set_sampling(1);
306 auto insert = samples.emplace(loc, LocalData{
307 threshold, // threshold
314 true // benching (if we have no data, we need at least one)
316 LocalData& data = insert.first->second;
318 XBT_DEBUG("XXXXX First time ever on benched nest %s.", loc.c_str());
319 xbt_assert(threshold > 0 || iters > 0,
320 "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
322 if (data.iters != iters || data.threshold != threshold) {
323 XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
324 "How did you manage to give two numbers at the same line??",
325 loc.c_str(), data.iters, data.threshold, iters, threshold);
329 // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate
330 // the computation instead
331 data.benching = data.need_more_benchs();
332 XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc.c_str(),
333 (data.benching ? "more benching needed" : "we have enough data, skip computes"));
337 int smpi_sample_2(int global, const char *file, int line)
339 SampleLocation loc(global, file, line);
342 XBT_DEBUG("sample2 %s", loc.c_str());
343 auto sample = samples.find(loc);
344 if (sample == samples.end())
345 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
346 LocalData& data = sample->second;
349 // we need to run a new bench
350 XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f",
351 data.count, data.iters, data.relstderr, data.threshold, data.mean);
354 // Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
355 //ran one bench and need to bail out now that our job is done). Just sleep instead
356 XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f)."
357 " apply the %fs delay instead",
358 data.count, data.iters, data.relstderr, data.threshold, data.mean);
359 smpi_execute(data.mean);
360 smpi_process()->set_sampling(0);
361 res = 0; // prepare to capture future, unrelated computations
367 void smpi_sample_3(int global, const char *file, int line)
369 SampleLocation loc(global, file, line);
371 XBT_DEBUG("sample3 %s", loc.c_str());
372 auto sample = samples.find(loc);
373 if (sample == samples.end())
374 xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
375 LocalData& data = sample->second;
377 if (not data.benching)
380 // ok, benchmarking this loop is over
381 xbt_os_threadtimer_stop(smpi_process()->timer());
385 double period = xbt_os_timer_elapsed(smpi_process()->timer());
387 data.sum_pow2 += period * period;
388 double n = static_cast<double>(data.count);
389 data.mean = data.sum / n;
390 data.relstderr = sqrt((data.sum_pow2 / n - data.mean * data.mean) / n) / data.mean;
391 if (data.need_more_benchs()) {
392 data.mean = period; // Still in benching process; We want sample_2 to simulate the exact time of this loop
393 // occurrence before leaving, not the mean over the history
395 XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)",
396 data.count, data.mean, data.relstderr, period);
398 // That's enough for now, prevent sample_2 to run the same code over and over
399 data.benching = false;
402 smpi_trace_call_location_t* smpi_trace_get_call_location()
404 return smpi_process()->call_location();
407 void smpi_trace_set_call_location(const char* file, const int line)
409 smpi_trace_call_location_t* loc = smpi_process()->call_location();
411 loc->previous_filename = loc->filename;
412 loc->previous_linenumber = loc->linenumber;
413 loc->filename = file;
414 loc->linenumber = line;
417 /** Required for Fortran bindings */
418 void smpi_trace_set_call_location_(const char* file, int* line)
420 smpi_trace_set_call_location(file, *line);
423 /** Required for Fortran if -fsecond-underscore is activated */
424 void smpi_trace_set_call_location__(const char* file, int* line)
426 smpi_trace_set_call_location(file, *line);
429 void smpi_bench_destroy()