X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/39c935d6d5ee86d153f6f7e6a10d723ae7c57f6f..321102577020b194dfb7ba89e48687952816849e:/src/smpi/internals/smpi_bench.cpp diff --git a/src/smpi/internals/smpi_bench.cpp b/src/smpi/internals/smpi_bench.cpp index 27956e9b3b..096049520e 100644 --- a/src/smpi/internals/smpi_bench.cpp +++ b/src/smpi/internals/smpi_bench.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2007-2021. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2022. 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. */ @@ -7,10 +7,13 @@ #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" @@ -34,8 +37,6 @@ static simgrid::config::Flag "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); @@ -56,7 +57,7 @@ void smpi_execute_flops(double 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, @@ -66,23 +67,19 @@ void smpi_execute(double 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; @@ -91,11 +88,9 @@ void smpi_bench_begin() 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()); @@ -132,14 +127,10 @@ void smpi_bench_end() int event_set = smpi_process()->papi_event_set(); std::vector 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 @@ -158,49 +149,45 @@ void smpi_bench_end() #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(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(), 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(private_sleep(usecs / 1000000.0)); + private_sleep(static_cast(usecs) / 1e6); + return 0; } #if _POSIX_TIMERS > 0 @@ -208,28 +195,26 @@ int smpi_nanosleep(const struct timespec* tp, struct timespec* t) { if (not smpi_process()) return nanosleep(tp,t); - return static_cast(private_sleep(tp->tv_sec + tp->tv_nsec / 1000000000.0)); + private_sleep(static_cast(tp->tv_sec) + static_cast(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(now); -#ifdef WIN32 - tv->tv_usec = static_cast((now - tv->tv_sec) * 1e6); -#else - tv->tv_usec = static_cast((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(secs); + tv->tv_usec = static_casttv_usec)>(usecs); // suseconds_t (or useconds_t on WIN32) } if (smpi_wtime_sleep > 0) simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep); - smpi_bench_begin(); return 0; } @@ -240,16 +225,17 @@ int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp) 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(now); - tp->tv_nsec = static_cast((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(secs); + tp->tv_nsec = static_cast(nsecs); if (smpi_wtime_sleep > 0) simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep); - smpi_bench_begin(); return 0; } #endif @@ -258,42 +244,34 @@ 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(); + 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(resolution); + const SmpiBenchGuard suspend_bench; + return static_cast(1.0 / sg_surf_precision); } unsigned long long smpi_rastro_timestamp () { - smpi_bench_end(); - double now = SIMIX_get_clock(); - - auto sec = static_cast(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(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())); @@ -316,9 +294,8 @@ public: 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; @@ -327,30 +304,31 @@ bool LocalData::need_more_benchs() const std::unordered_map> samples; } -void smpi_sample_1(int global, const char *file, int line, int iters, double threshold) +void smpi_sample_1(int global, const char *file, const char *tag, int iters, double threshold) { - 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) + }); + 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; + LocalData& data = sample->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??", @@ -366,14 +344,14 @@ void smpi_sample_1(int global, const char *file, int line, int iters, double thr } } -int smpi_sample_2(int global, const char *file, int line, int iter_count) +int smpi_sample_2(int global, const char *file,const char *tag, int iter_count) { - SampleLocation loc(global, file, line); + SampleLocation loc(global, file, tag); 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!"); + xbt_assert(sample != samples.end(), + "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!"); const LocalData& data = sample->second; if (data.benching) { @@ -385,10 +363,10 @@ int smpi_sample_2(int global, const char *file, int line, int iter_count) // 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= 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); @@ -402,14 +380,14 @@ int smpi_sample_2(int global, const char *file, int line, int iter_count) return 1; } -void smpi_sample_3(int global, const char *file, int line) +void smpi_sample_3(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()); 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) @@ -434,15 +412,15 @@ void smpi_sample_3(int global, const char *file, int line) 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);