X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/403af5e6247ce6452b721f418a5b41e4548efac4..a4c8f1898670317d0fa33bf1b1a904ea922b78cc:/src/smpi/internals/smpi_bench.cpp diff --git a/src/smpi/internals/smpi_bench.cpp b/src/smpi/internals/smpi_bench.cpp index 7f8dd61b15..efecf03056 100644 --- a/src/smpi/internals/smpi_bench.cpp +++ b/src/smpi/internals/smpi_bench.cpp @@ -1,14 +1,19 @@ -/* Copyright (c) 2007, 2009-2017. 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.h" +#include "src/mc/mc_replay.hpp" +#include "xbt/config.hpp" + +#include "src/smpi/include/smpi_actor.hpp" #include #ifndef WIN32 @@ -22,30 +27,26 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)"); +static simgrid::config::Flag + 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) { + 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); - simcall_set_category (action, TRACE_internal_smpi_get_category()); - simcall_execution_wait(action); - smpi_switch_data_segment(smpi_process()->index()); + 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(double duration) @@ -53,11 +54,8 @@ void smpi_execute(double duration) if (duration >= smpi_cpu_threshold) { XBT_DEBUG("Sleep for %g to handle real computation time", duration); double flops = duration * smpi_host_speed; - int rank = smpi_process()->index(); - instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); - extra->type=TRACING_COMPUTING; - extra->comp_size=flops; - TRACE_smpi_computing_in(rank, extra); + int rank = simgrid::s4u::this_actor::get_pid(); + TRACE_smpi_computing_in(rank, flops); smpi_execute_flops(flops); @@ -72,29 +70,28 @@ void smpi_execute(double duration) void smpi_execute_benched(double duration) { smpi_bench_end(); - smpi_execute(duration); + double speed = sg_host_speed(sg_host_self()); + smpi_execute_flops(duration*speed); smpi_bench_begin(); } void smpi_bench_begin() { - if (smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) { - smpi_switch_data_segment(smpi_process()->index()); + if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) { + smpi_switch_data_segment(simgrid::s4u::Actor::self()); } if (MC_is_active() || MC_record_replay_is_active()) return; #if HAVE_PAPI - if (not xbt_cfg_get_string("smpi/papi-events").empty()) { + if (not simgrid::config::get_value("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 @@ -115,7 +112,7 @@ 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 (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') { + if (not simgrid::config::get_value("smpi/papi-events").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()); @@ -138,8 +135,8 @@ void smpi_bench_end() 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. + // 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(); @@ -150,18 +147,19 @@ void smpi_bench_end() } // Simulate the benchmarked computation unless disabled via command-line argument - if (xbt_cfg_get_boolean("smpi/simulate-computation")) { + if (simgrid::config::get_value("smpi/simulate-computation")) { smpi_execute(xbt_os_timer_elapsed(timer)/speedup); } #if HAVE_PAPI - if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0' && TRACE_smpi_is_enabled()) { - container_t container = new simgrid::instr::Container(smpi_container(smpi_process()->index())); + if (not simgrid::config::get_value("smpi/papi-events").empty() && TRACE_smpi_is_enabled()) { + container_t 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(); 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(container->type_->by_name(pair.first)); + variable->set_event(SIMIX_get_clock(), pair.second); } } #endif @@ -169,17 +167,14 @@ void smpi_bench_end() 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(); - instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); - extra->type=TRACING_SLEEPING; - extra->sleep_duration=secs; - TRACE_smpi_sleeping_in(rank, extra); + int rank = simgrid::s4u::this_actor::get_pid(); + TRACE_smpi_sleeping_in(rank, secs); simcall_process_sleep(secs); @@ -191,23 +186,32 @@ static unsigned int private_sleep(double secs) unsigned int smpi_sleep(unsigned int secs) { - return private_sleep(static_cast(secs)); + if (not smpi_process()) + return sleep(secs); + return private_sleep(secs); } int smpi_usleep(useconds_t usecs) { - return static_cast(private_sleep(static_cast(usecs) / 1000000.0)); + if (not smpi_process()) + return usleep(usecs); + return static_cast(private_sleep(usecs / 1000000.0)); } #if _POSIX_TIMERS > 0 -int smpi_nanosleep(const struct timespec* tp, struct timespec* /*t*/) +int smpi_nanosleep(const struct timespec* tp, struct timespec* t) { - return static_cast(private_sleep(static_cast(tp->tv_sec + tp->tv_nsec / 1000000000.0))); + if (not smpi_process()) + return nanosleep(tp,t); + return static_cast(private_sleep(tp->tv_sec + tp->tv_nsec / 1000000000.0)); } #endif -int smpi_gettimeofday(struct timeval* tv, void* /*tz*/) +int smpi_gettimeofday(struct timeval* tv, struct timezone* tz) { + if (not smpi_process()) + return gettimeofday(tv, tz); + smpi_bench_end(); double now = SIMIX_get_clock(); if (tv) { @@ -218,13 +222,17 @@ int smpi_gettimeofday(struct timeval* tv, void* /*tz*/) tv->tv_usec = static_cast((now - tv->tv_sec) * 1e6); #endif } + if (smpi_wtime_sleep > 0) + simcall_process_sleep(smpi_wtime_sleep); smpi_bench_begin(); return 0; } #if _POSIX_TIMERS > 0 -int smpi_clock_gettime(clockid_t /*clk_id*/, struct timespec* tp) +int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp) { + 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(); @@ -232,11 +240,28 @@ int smpi_clock_gettime(clockid_t /*clk_id*/, struct timespec* tp) 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); 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) + simcall_process_sleep(smpi_wtime_sleep); + smpi_bench_begin(); + } else { + time = SIMIX_get_clock(); + } + return time; +} + extern double sg_surf_precision; unsigned long long smpi_rastro_resolution () { @@ -264,7 +289,7 @@ public: SampleLocation(bool global, const char* file, int line) : std::string(std::string(file) + ":" + std::to_string(line)) { if (not global) - this->append(":" + std::to_string(smpi_process()->index())); + this->append(":" + std::to_string(simgrid::s4u::this_actor::get_pid())); } }; @@ -281,9 +306,6 @@ public: bool need_more_benchs() const; }; -} - -std::unordered_map> samples; bool LocalData::need_more_benchs() const { @@ -295,12 +317,16 @@ bool LocalData::need_more_benchs() const return res; } +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 @@ -333,12 +359,11 @@ 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!"); @@ -346,21 +371,28 @@ int smpi_sample_2(int global, const char *file, int line) 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) @@ -384,13 +416,10 @@ 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); @@ -398,7 +427,25 @@ void smpi_sample_3(int global, const char *file, int line) data.benching = false; } -extern "C" { /** These functions will be called from the user code **/ +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(); @@ -425,9 +472,40 @@ void smpi_trace_set_call_location__(const char* file, int* line) { smpi_trace_set_call_location(file, *line); } -} void smpi_bench_destroy() { 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) +{ + if (smpi_process()) + optind = smpi_process()->get_optind(); + int ret = getopt_long (argc, argv, options, long_options, opt_index); + if (smpi_process()) + smpi_process()->set_optind(optind); + return ret; +} + +int smpi_getopt (int argc, char *const *argv, const char *options) +{ + if (smpi_process()) + optind = smpi_process()->get_optind(); + int ret = getopt (argc, argv, options); + if (smpi_process()) + smpi_process()->set_optind(optind); + return ret; +}