X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/31240e831c7cf5bc0769e411ab74018c2add3262..49e2303ab8fb6ac8f8f7a0fbbaa73528256396a7:/src/smpi/internals/smpi_bench.cpp diff --git a/src/smpi/internals/smpi_bench.cpp b/src/smpi/internals/smpi_bench.cpp index 4aa873a1f4..0c2d89b60e 100644 --- a/src/smpi/internals/smpi_bench.cpp +++ b/src/smpi/internals/smpi_bench.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2007-2019. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2020. 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. */ @@ -11,8 +11,8 @@ #include "smpi_comm.hpp" #include "src/internal_config.h" #include "src/mc/mc_replay.hpp" -#include "src/simix/ActorImpl.hpp" #include "xbt/config.hpp" +#include "xbt/file.hpp" #include "src/smpi/include/smpi_actor.hpp" #include @@ -20,6 +20,7 @@ #ifndef WIN32 #include #endif +#include #include #if HAVE_PAPI @@ -33,65 +34,53 @@ 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_flops(double flops) +{ + private_execute_flops(flops); +} + void smpi_execute(double duration) { - if (duration >= smpi_cpu_threshold) { + if (duration >= smpi_cfg_cpu_thresh()) { 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(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()); + 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(); + smpi_execute_flops(flops); + smpi_bench_begin(); +} + void smpi_bench_begin() { - if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) { + if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) { smpi_switch_data_segment(simgrid::s4u::Actor::self()); } @@ -99,7 +88,7 @@ void smpi_bench_begin() 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) { @@ -112,12 +101,24 @@ void smpi_bench_begin() 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); @@ -126,10 +127,10 @@ 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"); @@ -144,35 +145,25 @@ void smpi_bench_end() 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()) { + const 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)); + auto* variable = static_cast(container->type_->by_name(pair.first)); variable->set_event(SIMIX_get_clock(), pair.second); } } @@ -190,7 +181,7 @@ static unsigned int private_sleep(double secs) 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); @@ -202,14 +193,14 @@ unsigned int smpi_sleep(unsigned int secs) { if (not smpi_process()) return sleep(secs); - return private_sleep(static_cast(secs)); + return private_sleep(secs); } int smpi_usleep(useconds_t usecs) { if (not smpi_process()) return usleep(usecs); - return static_cast(private_sleep(static_cast(usecs) / 1000000.0)); + return static_cast(private_sleep(usecs / 1000000.0)); } #if _POSIX_TIMERS > 0 @@ -217,7 +208,7 @@ 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))); + return static_cast(private_sleep(tp->tv_sec + tp->tv_nsec / 1000000000.0)); } #endif @@ -237,7 +228,7 @@ int smpi_gettimeofday(struct timeval* tv, struct timezone* tz) #endif } if (smpi_wtime_sleep > 0) - simcall_process_sleep(smpi_wtime_sleep); + simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep); smpi_bench_begin(); return 0; } @@ -245,17 +236,19 @@ int smpi_gettimeofday(struct timeval* tv, struct timezone* tz) #if _POSIX_TIMERS > 0 int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp) { + if (not tp) { + errno = EFAULT; + return -1; + } if (not smpi_process()) 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); - } + tp->tv_sec = static_cast(now); + tp->tv_nsec = static_cast((now - tp->tv_sec) * 1e9); if (smpi_wtime_sleep > 0) - simcall_process_sleep(smpi_wtime_sleep); + simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep); smpi_bench_begin(); return 0; } @@ -268,7 +261,7 @@ double smpi_mpi_wtime() smpi_bench_end(); time = SIMIX_get_clock(); if (smpi_wtime_sleep > 0) - simcall_process_sleep(smpi_wtime_sleep); + simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep); smpi_bench_begin(); } else { time = SIMIX_get_clock(); @@ -290,10 +283,10 @@ unsigned long long smpi_rastro_timestamp () smpi_bench_end(); double now = SIMIX_get_clock(); - unsigned long long sec = static_cast(now); + auto sec = static_cast(now); unsigned long long pre = (now - sec) * smpi_rastro_resolution(); smpi_bench_begin(); - return static_cast(sec) * smpi_rastro_resolution() + pre; + return sec * smpi_rastro_resolution() + pre; } /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/ @@ -337,9 +330,10 @@ std::unordered_map> 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 @@ -351,12 +345,12 @@ void smpi_sample_1(int global, const char *file, int line, int iters, double thr 0, // count true // benching (if we have no data, we need at least one) }); - LocalData& data = insert.first->second; if (insert.second) { 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??", @@ -372,34 +366,40 @@ 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 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!"); - LocalData& data = sample->second; + const LocalData& data = sample->second; 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 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 >= iter (%d >= %d) or stderrset_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) @@ -423,13 +423,10 @@ 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); @@ -437,6 +434,25 @@ 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){ + 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!"); + + 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() { return smpi_process()->call_location(); @@ -448,18 +464,21 @@ 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); }