-/* Copyright (c) 2007-2018. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2007-2019. 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 "getopt.h"
#include "private.hpp"
#include "simgrid/host.h"
#include "simgrid/modelchecker.h"
+#include "simgrid/s4u/Exec.hpp"
#include "smpi_comm.hpp"
-#include "smpi_process.hpp"
#include "src/internal_config.h"
#include "src/mc/mc_replay.hpp"
-#include "src/simix/ActorImpl.hpp"
#include "xbt/config.hpp"
-#include "getopt.h"
+#include "src/smpi/include/smpi_actor.hpp"
#include <unordered_map>
#ifndef WIN32
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)");
+static simgrid::config::Flag<double>
+ smpi_wtime_sleep("smpi/wtime",
+ "Minimum time to inject inside a call to MPI_Wtime(), gettimeofday() and clock_gettime()",
+ 1e-8 /* Documented to be 10 ns */);
+
double smpi_cpu_threshold = -1;
double smpi_host_speed;
-shared_malloc_type smpi_cfg_shared_malloc = shmalloc_global;
+SharedMallocType smpi_cfg_shared_malloc = SharedMallocType::GLOBAL;
double smpi_total_benched_time = 0;
-extern "C" XBT_PUBLIC void smpi_execute_flops_(double* flops);
-void smpi_execute_flops_(double *flops)
-{
- smpi_execute_flops(*flops);
-}
-
-extern "C" XBT_PUBLIC void smpi_execute_(double* duration);
-void smpi_execute_(double *duration)
-{
- smpi_execute(*duration);
-}
-
-void smpi_execute_flops(double flops) {
+// Private execute_flops used by smpi_execute and spmi_execute_benched
+void private_execute_flops(double flops) {
xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
XBT_DEBUG("Handle real computation time: %f flops", flops);
- smx_activity_t action = simcall_execution_start("computation", flops, 1, 0, smpi_process()->process()->get_host());
- simcall_set_category (action, TRACE_internal_smpi_get_category());
- simcall_execution_wait(action);
+ simgrid::s4u::this_actor::exec_init(flops)
+ ->set_name("computation")
+ ->set_tracing_category(smpi_process()->get_tracing_category())
+ ->start()
+ ->wait();
smpi_switch_data_segment(simgrid::s4u::Actor::self());
}
+void smpi_execute_flops(double flops) {
+ int rank = simgrid::s4u::this_actor::get_pid();
+ TRACE_smpi_computing_in(rank, flops);
+
+ private_execute_flops(flops);
+
+ TRACE_smpi_computing_out(rank);
+}
+
void smpi_execute(double duration)
{
if (duration >= smpi_cpu_threshold) {
int rank = simgrid::s4u::this_actor::get_pid();
TRACE_smpi_computing_in(rank, flops);
- smpi_execute_flops(flops);
+ private_execute_flops(flops);
TRACE_smpi_computing_out(rank);
void smpi_bench_begin()
{
- if (smpi_privatize_global_variables == SmpiPrivStrategies::Mmap) {
+ if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
smpi_switch_data_segment(simgrid::s4u::Actor::self());
}
return;
#if HAVE_PAPI
- if (not xbt_cfg_get_string("smpi/papi-events").empty()) {
+ if (not simgrid::config::get_value<std::string>("smpi/papi-events").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()) {
- if (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() && PAPI_start(event_set) != PAPI_OK) {
+ // TODO This needs some proper handling.
+ XBT_CRITICAL("Could not start PAPI counters.\n");
+ xbt_die("Error.");
}
}
#endif
xbt_os_threadtimer_start(smpi_process()->timer());
}
+double smpi_adjust_comp_speed(){
+ double speedup=1;
+ if (simgrid::config::get_value<std::string>("smpi/comp-adjustment-file")[0] != '\0') {
+
+ smpi_trace_call_location_t* loc = smpi_process()->call_location();
+ std::string key = loc->get_composed_key();
+ std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
+ if (it != location2speedup.end()) {
+ speedup = it->second;
+ }
+ }
+ return speedup;
+}
+
void smpi_bench_end()
{
if (MC_is_active() || MC_record_replay_is_active())
return;
- double speedup = 1;
xbt_os_timer_t timer = smpi_process()->timer();
xbt_os_threadtimer_stop(timer);
* An MPI function has been called and now is the right time to update
* our PAPI counters for this process.
*/
- if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
+ if (not simgrid::config::get_value<std::string>("smpi/papi-events").empty()) {
papi_counter_t& counter_data = smpi_process()->papi_counters();
int event_set = smpi_process()->papi_event_set();
std::vector<long long> event_values = std::vector<long long>(counter_data.size());
xbt_die("Aborting.");
}
- if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') { // Maybe we need to artificially speed up or slow
- // down our computation based on our statistical analysis.
-
- smpi_trace_call_location_t* loc = smpi_process()->call_location();
- std::string key = loc->get_composed_key();
- std::unordered_map<std::string, double>::const_iterator it = location2speedup.find(key);
- if (it != location2speedup.end()) {
- speedup = it->second;
- }
- }
-
+ // Maybe we need to artificially speed up or slow down our computation based on our statistical analysis.
// Simulate the benchmarked computation unless disabled via command-line argument
- if (xbt_cfg_get_boolean("smpi/simulate-computation")) {
- smpi_execute(xbt_os_timer_elapsed(timer)/speedup);
+ if (simgrid::config::get_value<bool>("smpi/simulate-computation")) {
+ smpi_execute(xbt_os_timer_elapsed(timer)/smpi_adjust_comp_speed());
}
#if HAVE_PAPI
- if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0' && TRACE_smpi_is_enabled()) {
+ if (not simgrid::config::get_value<std::string>("smpi/papi-events").empty() && TRACE_smpi_is_enabled()) {
container_t container =
- new simgrid::instr::Container(std::string("rank-") + std::to_string(simgrid::s4u::this_actor::get_pid()));
+ simgrid::instr::Container::by_name(std::string("rank-") + std::to_string(simgrid::s4u::this_actor::get_pid()));
papi_counter_t& counter_data = smpi_process()->papi_counters();
for (auto const& pair : counter_data) {
- new simgrid::instr::SetVariableEvent(
- surf_get_clock(), container, PJ_type_get(/* countername */ pair.first.c_str(), container->type), pair.second);
+ simgrid::instr::VariableType* variable = static_cast<simgrid::instr::VariableType*>(container->type_->by_name(pair.first));
+ variable->set_event(SIMIX_get_clock(), pair.second);
}
}
#endif
smpi_total_benched_time += xbt_os_timer_elapsed(timer);
}
-/* Private sleep function used by smpi_sleep() and smpi_usleep() */
+/* Private sleep function used by smpi_sleep(), smpi_usleep() and friends */
static unsigned int private_sleep(double secs)
{
smpi_bench_end();
XBT_DEBUG("Sleep for: %lf secs", secs);
- int rank = MPI_COMM_WORLD->rank();
+ int rank = simgrid::s4u::this_actor::get_pid();
TRACE_smpi_sleeping_in(rank, secs);
- simcall_process_sleep(secs);
+ simgrid::s4u::this_actor::sleep_for(secs);
TRACE_smpi_sleeping_out(rank);
{
if (not smpi_process())
return sleep(secs);
- return private_sleep(static_cast<double>(secs));
+ return private_sleep(secs);
}
int smpi_usleep(useconds_t usecs)
{
if (not smpi_process())
return usleep(usecs);
- return static_cast<int>(private_sleep(static_cast<double>(usecs) / 1000000.0));
+ return static_cast<int>(private_sleep(usecs / 1000000.0));
}
#if _POSIX_TIMERS > 0
{
if (not smpi_process())
return nanosleep(tp,t);
- return static_cast<int>(private_sleep(static_cast<double>(tp->tv_sec + tp->tv_nsec / 1000000000.0)));
+ return static_cast<int>(private_sleep(tp->tv_sec + tp->tv_nsec / 1000000000.0));
}
#endif
tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
#endif
}
+ if (smpi_wtime_sleep > 0)
+ simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
smpi_bench_begin();
return 0;
}
tp->tv_sec = static_cast<time_t>(now);
tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
}
+ if (smpi_wtime_sleep > 0)
+ simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
smpi_bench_begin();
return 0;
}
#endif
+double smpi_mpi_wtime()
+{
+ double time;
+ if (smpi_process()->initialized() && not smpi_process()->finalized() && not smpi_process()->sampling()) {
+ smpi_bench_end();
+ time = SIMIX_get_clock();
+ if (smpi_wtime_sleep > 0)
+ simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep);
+ smpi_bench_begin();
+ } else {
+ time = SIMIX_get_clock();
+ }
+ return time;
+}
+
extern double sg_surf_precision;
unsigned long long smpi_rastro_resolution ()
{
bool need_more_benchs() const;
};
-}
-
-std::unordered_map<SampleLocation, LocalData, std::hash<std::string>> samples;
bool LocalData::need_more_benchs() const
{
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)
{
SampleLocation loc(global, file, line);
-
- smpi_bench_end(); /* Take time from previous, unrelated computation into account */
- smpi_process()->set_sampling(1);
+ 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
}
}
-int smpi_sample_2(int global, const char *file, int line)
+int smpi_sample_2(int global, const char *file, int line, int iter_count)
{
SampleLocation loc(global, file, line);
- int res;
- XBT_DEBUG("sample2 %s", loc.c_str());
+ 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!");
if (data.benching) {
// we need to run a new bench
- XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f",
- data.count, data.iters, data.relstderr, data.threshold, data.mean);
- res = 1;
+ XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f; total:%f",
+ data.count, data.iters, data.relstderr, data.threshold, data.mean, data.sum);
+ smpi_bench_begin();
} else {
// 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
- XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f)."
- " apply the %fs delay instead",
- data.count, data.iters, data.relstderr, data.threshold, data.mean);
- smpi_execute(data.mean);
- smpi_process()->set_sampling(0);
- res = 0; // prepare to capture future, unrelated computations
+ 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). "
+ "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);
+ smpi_bench_begin();
+ return 0;
+ } else {
+ XBT_DEBUG("Skipping - Benchmark already performed - accumulating time");
+ xbt_os_threadtimer_start(smpi_process()->timer());
+ }
}
- smpi_bench_begin();
- return res;
+ return 1;
}
void smpi_sample_3(int global, const char *file, int line)
double period = xbt_os_timer_elapsed(smpi_process()->timer());
data.sum += period;
data.sum_pow2 += period * period;
- double n = static_cast<double>(data.count);
+ double n = data.count;
data.mean = data.sum / n;
data.relstderr = sqrt((data.sum_pow2 / n - data.mean * data.mean) / n) / data.mean;
- if (data.need_more_benchs()) {
- data.mean = period; // Still in benching process; We want sample_2 to simulate the exact time of this loop
- // occurrence before leaving, not the mean over the history
- }
+
XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)",
data.count, data.mean, data.relstderr, period);
data.benching = false;
}
+int smpi_sample_exit(int global, const char *file, int line, int iter_count){
+ if (smpi_process()->sampling()){
+ SampleLocation loc(global, file, line);
+
+ 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!");
+ LocalData& data = sample->second;
+
+ if (smpi_process()->sampling()){//end of loop, but still sampling needed
+ smpi_process()->set_sampling(0);
+ smpi_execute(data.mean*iter_count);
+ smpi_bench_begin();
+ }
+ }
+ return 0;
+}
+
smpi_trace_call_location_t* smpi_trace_get_call_location()
{
return smpi_process()->call_location();
samples.clear();
}
+int smpi_getopt_long_only (int argc, char *const *argv, const char *options,
+ const struct option * long_options, int *opt_index)
+{
+ if (smpi_process())
+ optind = smpi_process()->get_optind();
+ int ret = getopt_long_only (argc, argv, options, long_options, opt_index);
+ if (smpi_process())
+ smpi_process()->set_optind(optind);
+ return ret;
+}
+
int smpi_getopt_long (int argc, char *const *argv, const char *options,
const struct option * long_options, int *opt_index)
{