X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/a1234c9c887de28ed21e81528886c20456c7b6cf..00f2021a5b4cf37f52dab7839d0d29505a60e673:/src/smpi/internals/smpi_bench.cpp diff --git a/src/smpi/internals/smpi_bench.cpp b/src/smpi/internals/smpi_bench.cpp index 9fbc6b0dc5..d79504f4f7 100644 --- a/src/smpi/internals/smpi_bench.cpp +++ b/src/smpi/internals/smpi_bench.cpp @@ -1,26 +1,28 @@ -/* Copyright (c) 2007-2019. 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 "getopt.h" #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/kernel/lmm/System.hpp" // sg_precision_timing #include "src/mc/mc_replay.hpp" -#include "src/simix/ActorImpl.hpp" #include "xbt/config.hpp" +#include "xbt/file.hpp" +#include #include "src/smpi/include/smpi_actor.hpp" -#include -#ifndef WIN32 -#include -#endif +#include #include +#include +#include #if HAVE_PAPI #include @@ -33,91 +35,83 @@ 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_cpu_threshold = -1; -double smpi_host_speed; - -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 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); XBT_DEBUG("Handle real computation time: %f flops", flops); simgrid::s4u::this_actor::exec_init(flops) ->set_name("computation") - ->set_tracing_category(TRACE_internal_smpi_get_category()) + ->set_tracing_category(smpi_process()->get_tracing_category()) ->start() ->wait(); smpi_switch_data_segment(simgrid::s4u::Actor::self()); } -void smpi_execute(double duration) +void smpi_execute_flops(double flops) { - if (duration >= smpi_cpu_threshold) { - XBT_DEBUG("Sleep for %g to handle real computation time", duration); - double flops = duration * smpi_host_speed; - int rank = simgrid::s4u::this_actor::get_pid(); - TRACE_smpi_computing_in(rank, flops); - - smpi_execute_flops(flops); - - TRACE_smpi_computing_out(rank); + private_execute_flops(flops); +} +void smpi_execute(double duration) +{ + if (duration >= smpi_cfg_cpu_thresh()) { + 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, - smpi_cpu_threshold); + smpi_cfg_cpu_thresh()); } } void smpi_execute_benched(double duration) { - smpi_bench_end(); - double speed = sg_host_speed(sg_host_self()); + 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) { + const SmpiBenchGuard suspend_bench; + smpi_execute_flops(flops); } void smpi_bench_begin() { - if (smpi_privatize_global_variables == 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 HAVE_PAPI - if (not simgrid::config::get_value("smpi/papi-events").empty()) { + 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()); } +double smpi_adjust_comp_speed(){ + double speedup=1; + if (smpi_cfg_comp_adjustment_file()[0] != '\0') { + const smpi_trace_call_location_t* loc = smpi_process()->call_location(); + std::string key = loc->get_composed_key(); + std::unordered_map::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); @@ -126,90 +120,72 @@ void smpi_bench_end() * An MPI function has been called and now is the right time to update * our PAPI counters for this process. */ - if (not simgrid::config::get_value("smpi/papi-events").empty()) { + if (not smpi_cfg_papi_events_file().empty()) { papi_counter_t& counter_data = smpi_process()->papi_counters(); int event_set = smpi_process()->papi_event_set(); - std::vector event_values = std::vector(counter_data.size()); + 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 if (smpi_process()->sampling()) { XBT_CRITICAL("Cannot do recursive benchmarks."); - XBT_CRITICAL("Are you trying to make a call to MPI within a SMPI_SAMPLE_ block?"); + XBT_CRITICAL("Are you trying to make a call to MPI within an SMPI_SAMPLE_ block?"); xbt_backtrace_display_current(); xbt_die("Aborting."); } // Maybe we need to artificially speed up or slow down our computation based on our statistical analysis. - if (simgrid::config::get_value("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::const_iterator it = location2speedup.find(key); - if (it != location2speedup.end()) { - speedup = it->second; - } - } - // Simulate the benchmarked computation unless disabled via command-line argument - if (simgrid::config::get_value("smpi/simulate-computation")) { - smpi_execute(xbt_os_timer_elapsed(timer)/speedup); + if (smpi_cfg_simulate_computation()) { + smpi_execute(xbt_os_timer_elapsed(timer)/smpi_adjust_comp_speed()); } #if HAVE_PAPI - if (not simgrid::config::get_value("smpi/papi-events").empty() && TRACE_smpi_is_enabled()) { - container_t container = + if (not smpi_cfg_papi_events_file().empty() && TRACE_smpi_is_enabled()) { + simgrid::instr::Container* container = 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(); + const papi_counter_t& counter_data = smpi_process()->papi_counters(); - for (auto const& pair : counter_data) { - simgrid::instr::VariableType* 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(), static_cast(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); - - simcall_process_sleep(secs); - - TRACE_smpi_sleeping_out(rank); - - smpi_bench_begin(); - return 0; + 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(pid); } unsigned int smpi_sleep(unsigned int secs) { if (not smpi_process()) return sleep(secs); - return private_sleep(static_cast(secs)); + private_sleep(secs); + return 0; } int smpi_usleep(useconds_t usecs) { if (not smpi_process()) return usleep(usecs); - return static_cast(private_sleep(static_cast(usecs) / 1000000.0)); + private_sleep(static_cast(usecs) / 1e6); + return 0; } #if _POSIX_TIMERS > 0 @@ -217,46 +193,47 @@ int smpi_nanosleep(const struct timespec* tp, struct timespec* t) { if (not smpi_process()) return nanosleep(tp,t); - return static_cast(private_sleep(static_cast(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 } if (smpi_wtime_sleep > 0) - simcall_process_sleep(smpi_wtime_sleep); - smpi_bench_begin(); + simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep); return 0; } #if _POSIX_TIMERS > 0 int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp) { - if (not smpi_process()) + if (not tp) { + errno = EFAULT; + return -1; + } + 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(); - if (tp) { - 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) - simcall_process_sleep(smpi_wtime_sleep); - smpi_bench_begin(); + simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep); return 0; } #endif @@ -265,42 +242,34 @@ 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) - simcall_process_sleep(smpi_wtime_sleep); - smpi_bench_begin(); + simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep); } 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_precision_timing); } unsigned long long smpi_rastro_timestamp () { - smpi_bench_end(); - double now = SIMIX_get_clock(); - - unsigned long long sec = static_cast(now); - unsigned long long pre = (now - sec) * smpi_rastro_resolution(); - smpi_bench_begin(); - return static_cast(sec) * smpi_rastro_resolution() + pre; + const SmpiBenchGuard suspend_bench; + return static_cast(simgrid::s4u::Engine::get_clock() / sg_precision_timing); } /* ****************************** 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())); @@ -323,9 +292,8 @@ 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; @@ -334,25 +302,28 @@ 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) +int smpi_sample_cond(int global, const char* file, const char* tag, int iters, double threshold, int iter_count) { - SampleLocation loc(global, file, line); - - 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) - }); - LocalData& data = insert.first->second; - if (insert.second) { + 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 [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)"); @@ -364,59 +335,50 @@ void smpi_sample_1(int global, const char *file, int line, int iters, double thr 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) -{ - SampleLocation loc(global, file, line); - int res; - - XBT_DEBUG("sample2 %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; + 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", - 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 - if (not data.need_more_benchs()) - XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderrset_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 (%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); + 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) +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()); @@ -426,18 +388,31 @@ 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(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); +} + +int smpi_sample_exit(int global, const char *file, const char* tag, int iter_count){ + if (smpi_process()->sampling()){ + SampleLocation loc(global, file, tag); + + XBT_DEBUG("sample exit %s", loc.c_str()); + auto sample = samples.find(loc); + xbt_assert(sample != samples.end(), + "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!"); - // That's enough for now, prevent sample_2 to run the same code over and over - data.benching = false; + if (smpi_process()->sampling()){//end of loop, but still sampling needed + const LocalData& data = sample->second; + 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() @@ -451,18 +426,21 @@ void smpi_trace_set_call_location(const char* file, const int line) loc->previous_filename = loc->filename; loc->previous_linenumber = loc->linenumber; - loc->filename = file; - loc->linenumber = line; + if(not smpi_cfg_trace_call_use_absolute_path()) + loc->filename = simgrid::xbt::Path(file).get_base_name(); + else + loc->filename = file; + loc->linenumber = line; } /** Required for Fortran bindings */ -void smpi_trace_set_call_location_(const char* file, int* line) +void smpi_trace_set_call_location_(const char* file, const int* line) { smpi_trace_set_call_location(file, *line); } /** Required for Fortran if -fsecond-underscore is activated */ -void smpi_trace_set_call_location__(const char* file, int* line) +void smpi_trace_set_call_location__(const char* file, const int* line) { smpi_trace_set_call_location(file, *line); } @@ -503,3 +481,7 @@ int smpi_getopt (int argc, char *const *argv, const char *options) smpi_process()->set_optind(optind); return ret; } + +pid_t smpi_getpid(){ + return static_cast(simgrid::s4u::this_actor::get_pid()); +}