X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/09212820ce67437a2ba2c7ce65c97cc0d721ca24..fb5dcd3589886a5f3845bbdd77fab5edbb82a842:/src/smpi/internals/smpi_bench.cpp diff --git a/src/smpi/internals/smpi_bench.cpp b/src/smpi/internals/smpi_bench.cpp index 8bf464a586..277da718d0 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) 2007-2021. 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. */ @@ -9,9 +9,11 @@ #include "simgrid/modelchecker.h" #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 "xbt/config.hpp" +#include "xbt/file.hpp" #include "src/smpi/include/smpi_actor.hpp" #include @@ -19,6 +21,7 @@ #ifndef WIN32 #include #endif +#include #include #if HAVE_PAPI @@ -32,13 +35,7 @@ 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; - -// Private execute_flops used by smpi_execute and spmi_execute benched +// 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); @@ -50,44 +47,39 @@ void private_execute_flops(double flops) { 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); - +void smpi_execute_flops(double flops) +{ private_execute_flops(flops); - - TRACE_smpi_computing_out(rank); } 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); - - private_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()); } @@ -95,25 +87,35 @@ 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) { - // 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); @@ -122,59 +124,44 @@ 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); + container->get_variable(pair.first)->set_event(SIMIX_get_clock(), pair.second); } } #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 */ @@ -183,12 +170,12 @@ static unsigned int private_sleep(double secs) smpi_bench_end(); 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); + TRACE_smpi_sleeping_out(pid); smpi_bench_begin(); return 0; @@ -219,7 +206,7 @@ int smpi_nanosleep(const struct timespec* tp, struct timespec* t) 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(); @@ -241,15 +228,17 @@ 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 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); - } + tp->tv_sec = static_cast(now); + tp->tv_nsec = static_cast((now - tp->tv_sec) * 1e9); if (smpi_wtime_sleep > 0) simgrid::s4u::this_actor::sleep_for(smpi_wtime_sleep); smpi_bench_begin(); @@ -286,10 +275,10 @@ 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 ************************************/ @@ -348,12 +337,12 @@ 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??", @@ -375,9 +364,9 @@ int smpi_sample_2(int global, const char *file, int line, int iter_count) 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; + 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) { // we need to run a new bench @@ -411,8 +400,8 @@ void smpi_sample_3(int global, const char *file, int line) 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) @@ -443,14 +432,14 @@ int smpi_sample_exit(int global, const char *file, int line, int iter_count){ 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; - + 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 - smpi_process()->set_sampling(0); - smpi_execute(data.mean*iter_count); - smpi_bench_begin(); + const LocalData& data = sample->second; + smpi_process()->set_sampling(0); + smpi_execute(data.mean * iter_count); + smpi_bench_begin(); } } return 0; @@ -467,18 +456,21 @@ 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); }