-/* Copyright (c) 2007-2021. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2007-2023. The SimGrid Team. All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
* under the terms of the license (GNU LGPL) which comes with this package. */
#include "private.hpp"
#include "simgrid/host.h"
#include "simgrid/modelchecker.h"
+#include "simgrid/s4u/Engine.hpp"
#include "simgrid/s4u/Exec.hpp"
#include "smpi_comm.hpp"
+#include "smpi_utils.hpp"
#include "src/internal_config.h"
#include "src/mc/mc_replay.hpp"
+#include "src/surf/surf_interface.hpp" // sg_surf_precision
#include "xbt/config.hpp"
#include "xbt/file.hpp"
#include "src/smpi/include/smpi_actor.hpp"
-#include <unordered_map>
-#ifndef WIN32
-#include <sys/mman.h>
-#endif
#include <cerrno>
#include <cmath>
+#include <sys/mman.h>
+#include <unordered_map>
#if HAVE_PAPI
#include <papi.h>
"Minimum time to inject inside a call to MPI_Wtime(), gettimeofday() and clock_gettime()",
1e-8 /* Documented to be 10 ns */);
-double smpi_total_benched_time = 0;
-
// Private execute_flops used by smpi_execute and smpi_execute_benched
void private_execute_flops(double flops) {
xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
void smpi_execute(double duration)
{
if (duration >= smpi_cfg_cpu_thresh()) {
- XBT_DEBUG("Sleep for %g to handle real computation time", duration);
+ XBT_DEBUG("Sleep for %gs (host time) to handle real computation time", duration);
private_execute_flops(duration * smpi_cfg_host_speed());
} else {
XBT_DEBUG("Real computation took %g while option smpi/cpu-threshold is set to %g => ignore it", duration,
void smpi_execute_benched(double duration)
{
- smpi_bench_end();
+ const SmpiBenchGuard suspend_bench;
double speed = sg_host_get_speed(sg_host_self());
smpi_execute_flops(duration*speed);
- smpi_bench_begin();
}
void smpi_execute_flops_benched(double flops) {
- smpi_bench_end();
+ const SmpiBenchGuard suspend_bench;
smpi_execute_flops(flops);
- smpi_bench_begin();
}
void smpi_bench_begin()
{
- if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
- smpi_switch_data_segment(simgrid::s4u::Actor::self());
- }
+ smpi_switch_data_segment(simgrid::s4u::Actor::self());
if (MC_is_active() || MC_record_replay_is_active())
return;
if (not smpi_cfg_papi_events_file().empty()) {
int event_set = smpi_process()->papi_event_set();
// PAPI_start sets everything to 0! See man(3) PAPI_start
- if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized() && PAPI_start(event_set) != PAPI_OK) {
- // TODO This needs some proper handling.
- XBT_CRITICAL("Could not start PAPI counters.\n");
- xbt_die("Error.");
- }
+ if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized() && event_set)
+ xbt_assert(PAPI_start(event_set) == PAPI_OK,
+ "Could not start PAPI counters (TODO: this needs some proper handling).");
}
#endif
xbt_os_threadtimer_start(smpi_process()->timer());
int event_set = smpi_process()->papi_event_set();
std::vector<long long> event_values(counter_data.size());
- if (PAPI_stop(event_set, &event_values[0]) != PAPI_OK) { // Error
- XBT_CRITICAL("Could not stop PAPI counters.\n");
- xbt_die("Error.");
- } else {
- for (unsigned int i = 0; i < counter_data.size(); i++) {
- counter_data[i].second += event_values[i];
- }
- }
+ if (event_set)
+ xbt_assert(PAPI_stop(event_set, &event_values[0]) == PAPI_OK, "Could not stop PAPI counters.");
+ for (unsigned int i = 0; i < counter_data.size(); i++)
+ counter_data[i].second += event_values[i];
}
#endif
#if HAVE_PAPI
if (not smpi_cfg_papi_events_file().empty() && TRACE_smpi_is_enabled()) {
- const simgrid::instr::Container* container =
+ simgrid::instr::Container* container =
simgrid::instr::Container::by_name(std::string("rank-") + std::to_string(simgrid::s4u::this_actor::get_pid()));
const papi_counter_t& counter_data = smpi_process()->papi_counters();
- for (auto const& pair : counter_data) {
- auto* variable = static_cast<simgrid::instr::VariableType*>(container->type_->by_name(pair.first));
- variable->set_event(SIMIX_get_clock(), pair.second);
+ for (auto const& [counter, value] : counter_data) {
+ container->get_variable(counter)->set_event(simgrid::s4u::Engine::get_clock(), static_cast<double>(value));
}
}
#endif
- smpi_total_benched_time += xbt_os_timer_elapsed(timer);
+ simgrid::smpi::utils::add_benched_time(xbt_os_timer_elapsed(timer));
}
/* Private sleep function used by smpi_sleep(), smpi_usleep() and friends */
-static unsigned int private_sleep(double secs)
+static void private_sleep(double secs)
{
- smpi_bench_end();
+ const SmpiBenchGuard suspend_bench;
XBT_DEBUG("Sleep for: %lf secs", secs);
- int rank = simgrid::s4u::this_actor::get_pid();
- TRACE_smpi_sleeping_in(rank, secs);
-
+ aid_t pid = simgrid::s4u::this_actor::get_pid();
+ TRACE_smpi_sleeping_in(pid, secs);
simgrid::s4u::this_actor::sleep_for(secs);
-
- TRACE_smpi_sleeping_out(rank);
-
- smpi_bench_begin();
- return 0;
+ TRACE_smpi_sleeping_out(pid);
}
unsigned int smpi_sleep(unsigned int secs)
{
if (not smpi_process())
return sleep(secs);
- return private_sleep(secs);
+ private_sleep(secs);
+ return 0;
}
int smpi_usleep(useconds_t usecs)
{
if (not smpi_process())
return usleep(usecs);
- return static_cast<int>(private_sleep(usecs / 1000000.0));
+ private_sleep(static_cast<double>(usecs) / 1e6);
+ return 0;
}
#if _POSIX_TIMERS > 0
{
if (not smpi_process())
return nanosleep(tp,t);
- return static_cast<int>(private_sleep(tp->tv_sec + tp->tv_nsec / 1000000000.0));
+ private_sleep(static_cast<double>(tp->tv_sec) + static_cast<double>(tp->tv_nsec) / 1e9);
+ return 0;
}
#endif
int smpi_gettimeofday(struct timeval* tv, struct timezone* tz)
{
- if (not smpi_process())
+ if (not smpi_process()->initialized() || smpi_process()->finalized() || smpi_process()->sampling())
return gettimeofday(tv, tz);
- smpi_bench_end();
- double now = SIMIX_get_clock();
+ const SmpiBenchGuard suspend_bench;
if (tv) {
- tv->tv_sec = static_cast<time_t>(now);
-#ifdef WIN32
- tv->tv_usec = static_cast<useconds_t>((now - tv->tv_sec) * 1e6);
-#else
- tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
-#endif
+ double now = simgrid::s4u::Engine::get_clock();
+ double secs = trunc(now);
+ double usecs = (now - secs) * 1e6;
+ tv->tv_sec = static_cast<time_t>(secs);
+ tv->tv_usec = static_cast<decltype(tv->tv_usec)>(usecs); // suseconds_t
}
if (smpi_wtime_sleep > 0)
simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
- smpi_bench_begin();
return 0;
}
errno = EFAULT;
return -1;
}
- if (not smpi_process())
+ if (not smpi_process()->initialized() || smpi_process()->finalized() || smpi_process()->sampling())
return clock_gettime(clk_id, tp);
//there is only one time in SMPI, so clk_id is ignored.
- smpi_bench_end();
- double now = SIMIX_get_clock();
- tp->tv_sec = static_cast<time_t>(now);
- tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
+ const SmpiBenchGuard suspend_bench;
+ double now = simgrid::s4u::Engine::get_clock();
+ double secs = trunc(now);
+ double nsecs = (now - secs) * 1e9;
+ tp->tv_sec = static_cast<time_t>(secs);
+ tp->tv_nsec = static_cast<long int>(nsecs);
if (smpi_wtime_sleep > 0)
simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
- smpi_bench_begin();
return 0;
}
#endif
{
double time;
if (smpi_process()->initialized() && not smpi_process()->finalized() && not smpi_process()->sampling()) {
- smpi_bench_end();
- time = SIMIX_get_clock();
+ const SmpiBenchGuard suspend_bench;
+ time = simgrid::s4u::Engine::get_clock();
if (smpi_wtime_sleep > 0)
simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
- smpi_bench_begin();
} else {
- time = SIMIX_get_clock();
+ time = simgrid::s4u::Engine::get_clock();
}
return time;
}
-extern double sg_surf_precision;
+// Used by Akypuera (https://github.com/schnorr/akypuera)
unsigned long long smpi_rastro_resolution ()
{
- smpi_bench_end();
- double resolution = (1/sg_surf_precision);
- smpi_bench_begin();
- return static_cast<unsigned long long>(resolution);
+ const SmpiBenchGuard suspend_bench;
+ return static_cast<unsigned long long>(1.0 / sg_surf_precision);
}
unsigned long long smpi_rastro_timestamp ()
{
- smpi_bench_end();
- double now = SIMIX_get_clock();
-
- auto sec = static_cast<unsigned long long>(now);
- unsigned long long pre = (now - sec) * smpi_rastro_resolution();
- smpi_bench_begin();
- return sec * smpi_rastro_resolution() + pre;
+ const SmpiBenchGuard suspend_bench;
+ return static_cast<unsigned long long>(simgrid::s4u::Engine::get_clock() / sg_surf_precision);
}
/* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
namespace {
class SampleLocation : public std::string {
public:
- SampleLocation(bool global, const char* file, int line) : std::string(std::string(file) + ":" + std::to_string(line))
+ SampleLocation(bool global, const char* file, const char* tag) : std::string(std::string(file) + ":" + std::string(tag))
{
if (not global)
this->append(":" + std::to_string(simgrid::s4u::this_actor::get_pid()));
bool LocalData::need_more_benchs() const
{
- bool res = (count < iters) || (threshold > 0.0 && (count < 2 || // not enough data
- relstderr > threshold // stderr too high yet
- ));
+ bool res = (count < iters) && (threshold < 0.0 || count < 2 || // not enough data
+ relstderr >= threshold); // stderr too high yet
XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
(res ? "need more data" : "enough benchs"), count, iters, relstderr, threshold, mean);
return res;
std::unordered_map<SampleLocation, LocalData, std::hash<std::string>> samples;
}
-void smpi_sample_1(int global, const char *file, int line, int iters, double threshold)
+int smpi_sample_cond(int global, const char* file, const char* tag, int iters, double threshold, int iter_count)
{
- SampleLocation loc(global, file, line);
+ SampleLocation loc(global, file, tag);
if (not smpi_process()->sampling()) { /* Only at first call when benchmarking, skip for next ones */
smpi_bench_end(); /* Take time from previous, unrelated computation into account */
smpi_process()->set_sampling(1);
}
- auto insert = samples.emplace(loc, LocalData{
- threshold, // threshold
- 0.0, // relstderr
- 0.0, // mean
- 0.0, // sum
- 0.0, // sum_pow2
- iters, // iters
- 0, // count
- true // benching (if we have no data, we need at least one)
- });
- if (insert.second) {
+ auto [sample, inserted] = samples.try_emplace(loc,
+ LocalData{
+ threshold, // threshold
+ 0.0, // relstderr
+ 0.0, // mean
+ 0.0, // sum
+ 0.0, // sum_pow2
+ iters, // iters
+ 0, // count
+ true // benching (if we have no data, we need at least one)
+ });
+ LocalData& data = sample->second;
+
+ if (inserted) {
XBT_DEBUG("XXXXX First time ever on benched nest %s.", loc.c_str());
xbt_assert(threshold > 0 || iters > 0,
"You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
} else {
- LocalData& data = insert.first->second;
if (data.iters != iters || data.threshold != threshold) {
XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. "
"How did you manage to give two numbers at the same line??",
THROW_IMPOSSIBLE;
}
- // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate
+ // if we already have some data, check whether we should get one more bench or whether we should emulate
// the computation instead
data.benching = data.need_more_benchs();
XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc.c_str(),
(data.benching ? "more benching needed" : "we have enough data, skip computes"));
}
-}
-
-int smpi_sample_2(int global, const char *file, int line, int iter_count)
-{
- SampleLocation loc(global, file, line);
-
- XBT_DEBUG("sample2 %s %d", loc.c_str(), iter_count);
- auto sample = samples.find(loc);
- if (sample == samples.end())
- xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
- const LocalData& data = sample->second;
+ XBT_DEBUG("sample cond %s %d", loc.c_str(), iter_count);
if (data.benching) {
// we need to run a new bench
XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f; total:%f",
// Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
//ran one bench and need to bail out now that our job is done). Just sleep instead
if (not data.need_more_benchs()){
- XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f). "
+ XBT_DEBUG("No benchmark (either no need, or just ran one): count (%d) >= iter (%d) (or <2) or stderr (%f) < thres (%f), or thresh is negative and ignored. "
"Mean is %f, will be injected %d times",
data.count, data.iters, data.relstderr, data.threshold, data.mean, iter_count);
-
+
//we ended benchmarking, let's inject all the time, now, and fast forward out of the loop.
smpi_process()->set_sampling(0);
smpi_execute(data.mean*iter_count);
return 1;
}
-void smpi_sample_3(int global, const char *file, int line)
+void smpi_sample_iter(int global, const char* file, const char* tag)
{
- SampleLocation loc(global, file, line);
+ SampleLocation loc(global, file, tag);
- XBT_DEBUG("sample3 %s", loc.c_str());
+ XBT_DEBUG("sample iter %s", loc.c_str());
auto sample = samples.find(loc);
- if (sample == samples.end())
- xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
+ xbt_assert(sample != samples.end(),
+ "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
LocalData& data = sample->second;
-
- if (not data.benching)
- THROW_IMPOSSIBLE;
+ xbt_assert(data.benching);
// ok, benchmarking this loop is over
xbt_os_threadtimer_stop(smpi_process()->timer());
XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)",
data.count, data.mean, data.relstderr, period);
-
- // That's enough for now, prevent sample_2 to run the same code over and over
- data.benching = false;
}
-int smpi_sample_exit(int global, const char *file, int line, int iter_count){
+int smpi_sample_exit(int global, const char *file, const char* tag, int iter_count){
if (smpi_process()->sampling()){
- SampleLocation loc(global, file, line);
+ SampleLocation loc(global, file, tag);
XBT_DEBUG("sample exit %s", loc.c_str());
auto sample = samples.find(loc);
- if (sample == samples.end())
- xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
-
+ xbt_assert(sample != samples.end(),
+ "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
+
if (smpi_process()->sampling()){//end of loop, but still sampling needed
const LocalData& data = sample->second;
smpi_process()->set_sampling(0);
smpi_process()->set_optind(optind);
return ret;
}
+
+pid_t smpi_getpid(){
+ return static_cast<pid_t>(simgrid::s4u::this_actor::get_pid());
+}
