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. */
8 #include <unordered_map>
11 #include "src/internal_config.h"
13 #include "private.hpp"
16 #include "xbt/sysdep.h"
18 #include "surf/surf.h"
19 #include "simgrid/sg_config.h"
20 #include "simgrid/modelchecker.h"
21 #include "src/mc/mc_replay.h"
23 #include <sys/types.h>
30 #include <math.h> // sqrt
39 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)");
42 xbt_dict_t samples = nullptr; /* Allocated on first use */
44 double smpi_cpu_threshold;
45 double smpi_host_speed;
47 shared_malloc_type smpi_cfg_shared_malloc = shmalloc_global;
48 double smpi_total_benched_time = 0;
49 smpi_privatisation_region_t smpi_privatisation_regions;
51 void smpi_bench_destroy()
53 xbt_dict_free(&samples);
56 extern "C" XBT_PUBLIC(void) smpi_execute_flops_(double *flops);
57 void smpi_execute_flops_(double *flops)
59 smpi_execute_flops(*flops);
62 extern "C" XBT_PUBLIC(void) smpi_execute_(double *duration);
63 void smpi_execute_(double *duration)
65 smpi_execute(*duration);
68 void smpi_execute_flops(double flops) {
69 XBT_DEBUG("Handle real computation time: %f flops", flops);
70 smx_activity_t action = simcall_execution_start("computation", flops, 1, 0);
71 simcall_set_category (action, TRACE_internal_smpi_get_category());
72 simcall_execution_wait(action);
73 smpi_switch_data_segment(smpi_process()->index());
76 void smpi_execute(double duration)
78 if (duration >= smpi_cpu_threshold) {
79 XBT_DEBUG("Sleep for %g to handle real computation time", duration);
80 double flops = duration * smpi_host_speed;
81 int rank = smpi_process()->index();
82 instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
83 extra->type=TRACING_COMPUTING;
84 extra->comp_size=flops;
85 TRACE_smpi_computing_in(rank, extra);
87 smpi_execute_flops(flops);
89 TRACE_smpi_computing_out(rank);
92 XBT_DEBUG("Real computation took %g while option smpi/cpu_threshold is set to %g => ignore it", duration,
97 void smpi_bench_begin()
99 if (smpi_privatize_global_variables) {
100 smpi_switch_data_segment(smpi_process()->index());
103 if (MC_is_active() || MC_record_replay_is_active())
107 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
108 int event_set = smpi_process()->papi_event_set();
109 // PAPI_start sets everything to 0! See man(3) PAPI_start
110 if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized()) {
111 if (PAPI_start(event_set) != PAPI_OK) {
112 // TODO This needs some proper handling.
113 XBT_CRITICAL("Could not start PAPI counters.\n");
119 xbt_os_threadtimer_start(smpi_process()->timer());
122 void smpi_bench_end()
124 if (MC_is_active() || MC_record_replay_is_active())
128 xbt_os_timer_t timer = smpi_process()->timer();
129 xbt_os_threadtimer_stop(timer);
133 * An MPI function has been called and now is the right time to update
134 * our PAPI counters for this process.
136 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
137 papi_counter_t& counter_data = smpi_process()->papi_counters();
138 int event_set = smpi_process()->papi_event_set();
139 std::vector<long long> event_values = std::vector<long long>(counter_data.size());
141 if (PAPI_stop(event_set, &event_values[0]) != PAPI_OK) { // Error
142 XBT_CRITICAL("Could not stop PAPI counters.\n");
145 for (unsigned int i = 0; i < counter_data.size(); i++) {
146 counter_data[i].second += event_values[i];
147 // XBT_DEBUG("[%i] PAPI: Counter %s: Value is now %lli (got increment by %lli\n", smpi_process()->index(),
148 // counter_data[i].first.c_str(), counter_data[i].second, event_values[i]);
154 if (smpi_process()->sampling()) {
155 XBT_CRITICAL("Cannot do recursive benchmarks.");
156 XBT_CRITICAL("Are you trying to make a call to MPI within a SMPI_SAMPLE_ block?");
157 xbt_backtrace_display_current();
158 xbt_die("Aborting.");
161 if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') { // Maybe we need to artificially speed up or slow
162 // down our computation based on our statistical analysis.
164 smpi_trace_call_location_t* loc = smpi_process()->call_location();
165 std::string key = loc->get_composed_key();
166 std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
167 if (it != location2speedup.end()) {
168 speedup = it->second;
172 // Simulate the benchmarked computation unless disabled via command-line argument
173 if (xbt_cfg_get_boolean("smpi/simulate-computation")) {
174 smpi_execute(xbt_os_timer_elapsed(timer)/speedup);
178 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0' && TRACE_smpi_is_enabled()) {
179 char container_name[INSTR_DEFAULT_STR_SIZE];
180 smpi_container(smpi_process()->index(), container_name, INSTR_DEFAULT_STR_SIZE);
181 container_t container = PJ_container_get(container_name);
182 papi_counter_t& counter_data = smpi_process()->papi_counters();
184 for (auto& pair : counter_data) {
185 new_pajeSetVariable(surf_get_clock(), container,
186 PJ_type_get(/* countername */ pair.first.c_str(), container->type), pair.second);
191 smpi_total_benched_time += xbt_os_timer_elapsed(timer);
194 /* Private sleep function used by smpi_sleep() and smpi_usleep() */
195 static unsigned int private_sleep(double secs)
199 XBT_DEBUG("Sleep for: %lf secs", secs);
200 int rank = MPI_COMM_WORLD->rank();
201 instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
202 extra->type=TRACING_SLEEPING;
203 extra->sleep_duration=secs;
204 TRACE_smpi_sleeping_in(rank, extra);
206 simcall_process_sleep(secs);
208 TRACE_smpi_sleeping_out(rank);
214 unsigned int smpi_sleep(unsigned int secs)
216 return private_sleep(static_cast<double>(secs));
219 int smpi_usleep(useconds_t usecs)
221 return static_cast<int>(private_sleep(static_cast<double>(usecs) / 1000000.0));
224 #if _POSIX_TIMERS > 0
225 int smpi_nanosleep(const struct timespec *tp, struct timespec * t)
227 return static_cast<int>(private_sleep(static_cast<double>(tp->tv_sec + tp->tv_nsec / 1000000000.0)));
231 int smpi_gettimeofday(struct timeval *tv, void* tz)
234 double now = SIMIX_get_clock();
236 tv->tv_sec = static_cast<time_t>(now);
238 tv->tv_usec = static_cast<useconds_t>((now - tv->tv_sec) * 1e6);
240 tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
247 #if _POSIX_TIMERS > 0
248 int smpi_clock_gettime(clockid_t clk_id, struct timespec *tp)
250 //there is only one time in SMPI, so clk_id is ignored.
252 double now = SIMIX_get_clock();
254 tp->tv_sec = static_cast<time_t>(now);
255 tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
262 extern double sg_surf_precision;
263 unsigned long long smpi_rastro_resolution ()
266 double resolution = (1/sg_surf_precision);
268 return static_cast<unsigned long long>(resolution);
271 unsigned long long smpi_rastro_timestamp ()
274 double now = SIMIX_get_clock();
276 unsigned long long sec = static_cast<unsigned long long>(now);
277 unsigned long long pre = (now - sec) * smpi_rastro_resolution();
279 return static_cast<unsigned long long>(sec) * smpi_rastro_resolution() + pre;
282 /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
284 double threshold; /* maximal stderr requested (if positive) */
285 double relstderr; /* observed stderr so far */
286 double mean; /* mean of benched times, to be used if the block is disabled */
287 double sum; /* sum of benched times (to compute the mean and stderr) */
288 double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
289 int iters; /* amount of requested iterations */
290 int count; /* amount of iterations done so far */
291 int benching; /* 1: we are benchmarking; 0: we have enough data, no bench anymore */
294 static char *sample_location(int global, const char *file, int line) {
296 return bprintf("%s:%d", file, line);
298 return bprintf("%s:%d:%d", file, line, smpi_process()->index());
302 static int sample_enough_benchs(local_data_t *data) {
303 int res = data->count >= data->iters;
304 if (data->threshold>0.0) {
306 res = 0; // not enough data
307 if (data->relstderr > data->threshold)
308 res = 0; // stderr too high yet
310 XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
311 (res?"enough benchs":"need more data"), data->count, data->iters, data->relstderr, data->threshold, data->mean);
315 void smpi_sample_1(int global, const char *file, int line, int iters, double threshold)
317 char *loc = sample_location(global, file, line);
319 smpi_bench_end(); /* Take time from previous, unrelated computation into account */
320 smpi_process()->set_sampling(1);
322 if (samples==nullptr)
323 samples = xbt_dict_new_homogeneous(free);
325 local_data_t *data = static_cast<local_data_t *>(xbt_dict_get_or_null(samples, loc));
327 xbt_assert(threshold>0 || iters>0,
328 "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
329 data = static_cast<local_data_t *>( xbt_new(local_data_t, 1));
332 data->sum_pow2 = 0.0;
334 data->threshold = threshold;
335 data->benching = 1; // If we have no data, we need at least one
337 xbt_dict_set(samples, loc, data, nullptr);
338 XBT_DEBUG("XXXXX First time ever on benched nest %s.",loc);
340 if (data->iters != iters || data->threshold != threshold) {
341 XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
342 "How did you manage to give two numbers at the same line??",
343 loc, data->iters, data->threshold, iters, threshold);
347 // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate
348 // the computation instead
349 data->benching = (sample_enough_benchs(data) == 0);
350 XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc,
351 (data->benching ? "more benching needed" : "we have enough data, skip computes"));
356 int smpi_sample_2(int global, const char *file, int line)
358 char *loc = sample_location(global, file, line);
361 xbt_assert(samples, "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
362 local_data_t *data = static_cast<local_data_t *>(xbt_dict_get(samples, loc));
363 XBT_DEBUG("sample2 %s",loc);
366 if (data->benching==1) {
367 // we need to run a new bench
368 XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f",
369 data->count, data->iters, data->relstderr, data->threshold, data->mean);
372 // Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
373 //ran one bench and need to bail out now that our job is done). Just sleep instead
374 XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f)."
375 " apply the %fs delay instead",
376 data->count, data->iters, data->relstderr, data->threshold, data->mean);
377 smpi_execute(data->mean);
378 smpi_process()->set_sampling(0);
379 res = 0; // prepare to capture future, unrelated computations
385 void smpi_sample_3(int global, const char *file, int line)
387 char *loc = sample_location(global, file, line);
389 xbt_assert(samples, "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
390 local_data_t *data = static_cast<local_data_t *>(xbt_dict_get(samples, loc));
391 XBT_DEBUG("sample3 %s",loc);
394 if (data->benching==0)
397 // ok, benchmarking this loop is over
398 xbt_os_threadtimer_stop(smpi_process()->timer());
402 double sample = xbt_os_timer_elapsed(smpi_process()->timer());
404 data->sum_pow2 += sample * sample;
405 double n = static_cast<double>(data->count);
406 data->mean = data->sum / n;
407 data->relstderr = sqrt((data->sum_pow2 / n - data->mean * data->mean) / n) / data->mean;
408 if (sample_enough_benchs(data)==0) {
409 data->mean = sample; // Still in benching process; We want sample_2 to simulate the exact time of this loop
410 // occurrence before leaving, not the mean over the history
412 XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)", data->count,
413 data->mean, data->relstderr, sample);
415 // That's enough for now, prevent sample_2 to run the same code over and over
419 extern "C" { /** These functions will be called from the user code **/
420 smpi_trace_call_location_t* smpi_trace_get_call_location() {
421 return smpi_process()->call_location();
424 void smpi_trace_set_call_location(const char* file, const int line) {
425 smpi_trace_call_location_t* loc = smpi_process()->call_location();
427 loc->previous_filename = loc->filename;
428 loc->previous_linenumber = loc->linenumber;
429 loc->filename = file;
430 loc->linenumber = line;
434 * Required for Fortran bindings
436 void smpi_trace_set_call_location_(const char* file, int* line) {
437 smpi_trace_set_call_location(file, *line);
441 * Required for Fortran if -fsecond-underscore is activated
443 void smpi_trace_set_call_location__(const char* file, int* line) {
444 smpi_trace_set_call_location(file, *line);