1 /* Copyright (c) 2007, 2009-2017. 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. */
6 #include "src/internal_config.h"
9 #include "simgrid/modelchecker.h"
10 #include "src/mc/mc_replay.h"
11 #include "src/smpi/smpi_process.hpp"
12 #include "src/smpi/smpi_comm.hpp"
17 #include <math.h> // sqrt
23 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)");
26 xbt_dict_t samples = nullptr; /* Allocated on first use */
28 double smpi_cpu_threshold = -1;
29 double smpi_host_speed;
31 shared_malloc_type smpi_cfg_shared_malloc = shmalloc_global;
32 double smpi_total_benched_time = 0;
33 smpi_privatisation_region_t smpi_privatisation_regions;
35 void smpi_bench_destroy()
37 xbt_dict_free(&samples);
40 extern "C" XBT_PUBLIC(void) smpi_execute_flops_(double *flops);
41 void smpi_execute_flops_(double *flops)
43 smpi_execute_flops(*flops);
46 extern "C" XBT_PUBLIC(void) smpi_execute_(double *duration);
47 void smpi_execute_(double *duration)
49 smpi_execute(*duration);
52 void smpi_execute_flops(double flops) {
53 XBT_DEBUG("Handle real computation time: %f flops", flops);
54 smx_activity_t action = simcall_execution_start("computation", flops, 1, 0);
55 simcall_set_category (action, TRACE_internal_smpi_get_category());
56 simcall_execution_wait(action);
57 smpi_switch_data_segment(smpi_process()->index());
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 = smpi_process()->index();
66 instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
67 extra->type=TRACING_COMPUTING;
68 extra->comp_size=flops;
69 TRACE_smpi_computing_in(rank, extra);
71 smpi_execute_flops(flops);
73 TRACE_smpi_computing_out(rank);
76 XBT_DEBUG("Real computation took %g while option smpi/cpu_threshold is set to %g => ignore it", duration,
81 void smpi_execute_benched(double duration)
84 smpi_execute(duration);
88 void smpi_bench_begin()
90 if (smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) {
91 smpi_switch_data_segment(smpi_process()->index());
94 if (MC_is_active() || MC_record_replay_is_active())
98 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
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 (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
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];
138 // XBT_DEBUG("[%i] PAPI: Counter %s: Value is now %lli (got increment by %lli\n", smpi_process()->index(),
139 // counter_data[i].first.c_str(), counter_data[i].second, event_values[i]);
145 if (smpi_process()->sampling()) {
146 XBT_CRITICAL("Cannot do recursive benchmarks.");
147 XBT_CRITICAL("Are you trying to make a call to MPI within a SMPI_SAMPLE_ block?");
148 xbt_backtrace_display_current();
149 xbt_die("Aborting.");
152 if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') { // Maybe we need to artificially speed up or slow
153 // down our computation based on our statistical analysis.
155 smpi_trace_call_location_t* loc = smpi_process()->call_location();
156 std::string key = loc->get_composed_key();
157 std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
158 if (it != location2speedup.end()) {
159 speedup = it->second;
163 // Simulate the benchmarked computation unless disabled via command-line argument
164 if (xbt_cfg_get_boolean("smpi/simulate-computation")) {
165 smpi_execute(xbt_os_timer_elapsed(timer)/speedup);
169 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0' && TRACE_smpi_is_enabled()) {
170 char container_name[INSTR_DEFAULT_STR_SIZE];
171 smpi_container(smpi_process()->index(), container_name, INSTR_DEFAULT_STR_SIZE);
172 container_t container = PJ_container_get(container_name);
173 papi_counter_t& counter_data = smpi_process()->papi_counters();
175 for (auto& pair : counter_data) {
176 new_pajeSetVariable(surf_get_clock(), container,
177 PJ_type_get(/* countername */ pair.first.c_str(), container->type), pair.second);
182 smpi_total_benched_time += xbt_os_timer_elapsed(timer);
185 /* Private sleep function used by smpi_sleep() and smpi_usleep() */
186 static unsigned int private_sleep(double secs)
190 XBT_DEBUG("Sleep for: %lf secs", secs);
191 int rank = MPI_COMM_WORLD->rank();
192 instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
193 extra->type=TRACING_SLEEPING;
194 extra->sleep_duration=secs;
195 TRACE_smpi_sleeping_in(rank, extra);
197 simcall_process_sleep(secs);
199 TRACE_smpi_sleeping_out(rank);
205 unsigned int smpi_sleep(unsigned int secs)
207 return private_sleep(static_cast<double>(secs));
210 int smpi_usleep(useconds_t usecs)
212 return static_cast<int>(private_sleep(static_cast<double>(usecs) / 1000000.0));
215 #if _POSIX_TIMERS > 0
216 int smpi_nanosleep(const struct timespec *tp, struct timespec * 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, void* tz)
225 double now = SIMIX_get_clock();
227 tv->tv_sec = static_cast<time_t>(now);
229 tv->tv_usec = static_cast<useconds_t>((now - tv->tv_sec) * 1e6);
231 tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
238 #if _POSIX_TIMERS > 0
239 int smpi_clock_gettime(clockid_t clk_id, struct timespec *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 double threshold; /* maximal stderr requested (if positive) */
276 double relstderr; /* observed stderr so far */
277 double mean; /* mean of benched times, to be used if the block is disabled */
278 double sum; /* sum of benched times (to compute the mean and stderr) */
279 double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
280 int iters; /* amount of requested iterations */
281 int count; /* amount of iterations done so far */
282 int benching; /* 1: we are benchmarking; 0: we have enough data, no bench anymore */
285 static char *sample_location(int global, const char *file, int line) {
287 return bprintf("%s:%d", file, line);
289 return bprintf("%s:%d:%d", file, line, smpi_process()->index());
293 static int sample_enough_benchs(local_data_t *data) {
294 int res = data->count >= data->iters;
295 if (data->threshold>0.0) {
297 res = 0; // not enough data
298 if (data->relstderr > data->threshold)
299 res = 0; // stderr too high yet
301 XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
302 (res?"enough benchs":"need more data"), data->count, data->iters, data->relstderr, data->threshold, data->mean);
306 void smpi_sample_1(int global, const char *file, int line, int iters, double threshold)
308 char *loc = sample_location(global, file, line);
310 smpi_bench_end(); /* Take time from previous, unrelated computation into account */
311 smpi_process()->set_sampling(1);
313 if (samples==nullptr)
314 samples = xbt_dict_new_homogeneous(free);
316 local_data_t *data = static_cast<local_data_t *>(xbt_dict_get_or_null(samples, loc));
318 xbt_assert(threshold>0 || iters>0,
319 "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
320 data = static_cast<local_data_t *>( xbt_new(local_data_t, 1));
323 data->sum_pow2 = 0.0;
325 data->threshold = threshold;
326 data->benching = 1; // If we have no data, we need at least one
328 xbt_dict_set(samples, loc, data, nullptr);
329 XBT_DEBUG("XXXXX First time ever on benched nest %s.",loc);
331 if (data->iters != iters || data->threshold != threshold) {
332 XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
333 "How did you manage to give two numbers at the same line??",
334 loc, data->iters, data->threshold, iters, threshold);
338 // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate
339 // the computation instead
340 data->benching = (sample_enough_benchs(data) == 0);
341 XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc,
342 (data->benching ? "more benching needed" : "we have enough data, skip computes"));
347 int smpi_sample_2(int global, const char *file, int line)
349 char *loc = sample_location(global, file, line);
352 xbt_assert(samples, "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
353 local_data_t *data = static_cast<local_data_t *>(xbt_dict_get(samples, loc));
354 XBT_DEBUG("sample2 %s",loc);
357 if (data->benching==1) {
358 // we need to run a new bench
359 XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f",
360 data->count, data->iters, data->relstderr, data->threshold, data->mean);
363 // Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
364 //ran one bench and need to bail out now that our job is done). Just sleep instead
365 XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f)."
366 " apply the %fs delay instead",
367 data->count, data->iters, data->relstderr, data->threshold, data->mean);
368 smpi_execute(data->mean);
369 smpi_process()->set_sampling(0);
370 res = 0; // prepare to capture future, unrelated computations
376 void smpi_sample_3(int global, const char *file, int line)
378 char *loc = sample_location(global, file, line);
380 xbt_assert(samples, "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
381 local_data_t *data = static_cast<local_data_t *>(xbt_dict_get(samples, loc));
382 XBT_DEBUG("sample3 %s",loc);
385 if (data->benching==0)
388 // ok, benchmarking this loop is over
389 xbt_os_threadtimer_stop(smpi_process()->timer());
393 double sample = xbt_os_timer_elapsed(smpi_process()->timer());
395 data->sum_pow2 += sample * sample;
396 double n = static_cast<double>(data->count);
397 data->mean = data->sum / n;
398 data->relstderr = sqrt((data->sum_pow2 / n - data->mean * data->mean) / n) / data->mean;
399 if (sample_enough_benchs(data)==0) {
400 data->mean = sample; // Still in benching process; We want sample_2 to simulate the exact time of this loop
401 // occurrence before leaving, not the mean over the history
403 XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)", data->count,
404 data->mean, data->relstderr, sample);
406 // That's enough for now, prevent sample_2 to run the same code over and over
410 extern "C" { /** These functions will be called from the user code **/
411 smpi_trace_call_location_t* smpi_trace_get_call_location() {
412 return smpi_process()->call_location();
415 void smpi_trace_set_call_location(const char* file, const int line) {
416 smpi_trace_call_location_t* loc = smpi_process()->call_location();
418 loc->previous_filename = loc->filename;
419 loc->previous_linenumber = loc->linenumber;
420 loc->filename = file;
421 loc->linenumber = line;
425 * Required for Fortran bindings
427 void smpi_trace_set_call_location_(const char* file, int* line) {
428 smpi_trace_set_call_location(file, *line);
432 * Required for Fortran if -fsecond-underscore is activated
434 void smpi_trace_set_call_location__(const char* file, int* line) {
435 smpi_trace_set_call_location(file, *line);