X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/f5792a3bf76ce15a573ae5e9c63097595ae5f2bd..54cb3695a30bf100bec2710d11af87d3b6ba1a44:/src/smpi/internals/smpi_bench.cpp diff --git a/src/smpi/internals/smpi_bench.cpp b/src/smpi/internals/smpi_bench.cpp index cbe42c8ee4..9793546b0f 100644 --- a/src/smpi/internals/smpi_bench.cpp +++ b/src/smpi/internals/smpi_bench.cpp @@ -1,15 +1,20 @@ -/* Copyright (c) 2007, 2009-2017. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2018. 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 "private.h" +#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 "src/simix/ActorImpl.hpp" +#include "xbt/config.hpp" + +#include "src/smpi/include/smpi_actor.hpp" #include #ifndef WIN32 @@ -23,31 +28,39 @@ 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; -smpi_privatization_region_t smpi_privatization_regions; -extern "C" XBT_PUBLIC(void) smpi_execute_flops_(double *flops); +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); +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(TRACE_internal_smpi_get_category()) + ->start() + ->wait(); + smpi_switch_data_segment(simgrid::s4u::Actor::self()); } void smpi_execute(double duration) @@ -55,11 +68,8 @@ 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); @@ -74,21 +84,22 @@ 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 (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') { + 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()) { @@ -117,7 +128,7 @@ 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()); @@ -128,8 +139,6 @@ void smpi_bench_end() } else { for (unsigned int i = 0; i < counter_data.size(); i++) { counter_data[i].second += event_values[i]; - // XBT_DEBUG("[%i] PAPI: Counter %s: Value is now %lli (got increment by %lli\n", smpi_process()->index(), - // counter_data[i].first.c_str(), counter_data[i].second, event_values[i]); } } } @@ -142,8 +151,8 @@ 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(); @@ -154,20 +163,19 @@ 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()) { - char container_name[INSTR_DEFAULT_STR_SIZE]; - smpi_container(smpi_process()->index(), container_name, INSTR_DEFAULT_STR_SIZE); - container_t container = PJ_container_get(container_name); + 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& pair : counter_data) { - new_pajeSetVariable(surf_get_clock(), container, - PJ_type_get(/* countername */ pair.first.c_str(), container->type), pair.second); + 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); } } #endif @@ -175,17 +183,14 @@ 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); @@ -197,23 +202,32 @@ static unsigned int private_sleep(double secs) unsigned int smpi_sleep(unsigned int secs) { + if (not smpi_process()) + return sleep(secs); return private_sleep(static_cast(secs)); } int smpi_usleep(useconds_t usecs) { + if (not smpi_process()) + return usleep(usecs); return static_cast(private_sleep(static_cast(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) { + if (not smpi_process()) + return nanosleep(tp,t); return static_cast(private_sleep(static_cast(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) { @@ -224,13 +238,17 @@ 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(); @@ -238,11 +256,28 @@ 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 +299,18 @@ unsigned long long smpi_rastro_timestamp () } /* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/ -typedef struct { +namespace { +class SampleLocation : public std::string { +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(simgrid::s4u::this_actor::get_pid())); + } +}; + +class LocalData { +public: double threshold; /* maximal stderr requested (if positive) */ double relstderr; /* observed stderr so far */ double mean; /* mean of benched times, to be used if the block is disabled */ @@ -272,95 +318,85 @@ typedef struct { double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */ int iters; /* amount of requested iterations */ int count; /* amount of iterations done so far */ - int benching; /* 1: we are benchmarking; 0: we have enough data, no bench anymore */ -} local_data_t; - -std::unordered_map samples; /* Allocated on first use */ + bool benching; /* true: we are benchmarking; false: we have enough data, no bench anymore */ -static char *sample_location(int global, const char *file, int line) { - if (global) { - return bprintf("%s:%d", file, line); - } else { - return bprintf("%s:%d:%d", file, line, smpi_process()->index()); - } -} + bool need_more_benchs() const; +}; -static int sample_enough_benchs(local_data_t *data) { - int res = data->count >= data->iters; - if (data->threshold>0.0) { - if (data->count <2) - res = 0; // not enough data - if (data->relstderr > data->threshold) - res = 0; // stderr too high yet - } +bool LocalData::need_more_benchs() const +{ + 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?"enough benchs":"need more data"), data->count, data->iters, data->relstderr, data->threshold, data->mean); + (res ? "need more data" : "enough benchs"), count, iters, relstderr, threshold, mean); return res; } +std::unordered_map> samples; +} + void smpi_sample_1(int global, const char *file, int line, int iters, double threshold) { - char *loc = sample_location(global, file, line); + SampleLocation loc(global, file, line); smpi_bench_end(); /* Take time from previous, unrelated computation into account */ smpi_process()->set_sampling(1); - auto ld = samples.find(loc); - local_data_t* data; - if (ld == samples.end()) { - xbt_assert(threshold>0 || iters>0, + 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) { + 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)"); - data = static_cast(xbt_new(local_data_t, 1)); - data->count = 0; - data->sum = 0.0; - data->sum_pow2 = 0.0; - data->iters = iters; - data->threshold = threshold; - data->benching = 1; // If we have no data, we need at least one - data->mean = 0; - samples[loc] = data; - XBT_DEBUG("XXXXX First time ever on benched nest %s.",loc); } else { - data = ld->second; - if (data->iters != iters || data->threshold != threshold) { + 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??", - loc, data->iters, data->threshold, iters, threshold); + loc.c_str(), data.iters, data.threshold, iters, threshold); THROW_IMPOSSIBLE; } // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate // the computation instead - data->benching = (sample_enough_benchs(data) == 0); - XBT_DEBUG("XXXX Re-entering the benched nest %s. %s", loc, - (data->benching ? "more benching needed" : "we have enough data, skip computes")); + 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")); } - xbt_free(loc); } int smpi_sample_2(int global, const char *file, int line) { - char *loc = sample_location(global, file, line); + SampleLocation loc(global, file, line); int res; - xbt_assert(not samples.empty(), - "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!"); - local_data_t* data = samples.at(loc); - XBT_DEBUG("sample2 %s",loc); - xbt_free(loc); + 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; - if (data->benching==1) { + 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); + data.count, data.iters, data.relstderr, data.threshold, data.mean); res = 1; } 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 stderrcount, data->iters, data->relstderr, data->threshold, data->mean); - smpi_execute(data->mean); + data.count, data.iters, data.relstderr, data.threshold, data.mean); + smpi_execute(data.mean); smpi_process()->set_sampling(0); res = 0; // prepare to capture future, unrelated computations } @@ -370,40 +406,39 @@ int smpi_sample_2(int global, const char *file, int line) void smpi_sample_3(int global, const char *file, int line) { - char *loc = sample_location(global, file, line); + SampleLocation loc(global, file, line); - xbt_assert(not samples.empty(), - "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!"); - local_data_t* data = samples.at(loc); - XBT_DEBUG("sample3 %s",loc); - xbt_free(loc); + 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!"); + LocalData& data = sample->second; - if (data->benching==0) + if (not data.benching) THROW_IMPOSSIBLE; // ok, benchmarking this loop is over xbt_os_threadtimer_stop(smpi_process()->timer()); // update the stats - data->count++; - double sample = xbt_os_timer_elapsed(smpi_process()->timer()); - data->sum += sample; - data->sum_pow2 += sample * sample; - double n = static_cast(data->count); - data->mean = data->sum / n; - data->relstderr = sqrt((data->sum_pow2 / n - data->mean * data->mean) / n) / data->mean; - if (sample_enough_benchs(data)==0) { - data->mean = sample; // Still in benching process; We want sample_2 to simulate the exact time of this loop + data.count++; + double period = xbt_os_timer_elapsed(smpi_process()->timer()); + data.sum += period; + data.sum_pow2 += period * period; + double n = static_cast(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, sample); + XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)", + data.count, data.mean, data.relstderr, period); // That's enough for now, prevent sample_2 to run the same code over and over - data->benching = 0; + data.benching = false; } -extern "C" { /** These functions will be called from the user code **/ smpi_trace_call_location_t* smpi_trace_get_call_location() { return smpi_process()->call_location(); @@ -430,10 +465,40 @@ void smpi_trace_set_call_location__(const char* file, int* line) { smpi_trace_set_call_location(file, *line); } -} void smpi_bench_destroy() { - for (auto elm : samples) - xbt_free(elm.second); + 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; }