X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/6a8d5cf4199454b5f811043c207d195107743e72..b9945098e2ba4e7b639e0c2245835d44800c1b29:/src/smpi/internals/smpi_bench.cpp diff --git a/src/smpi/internals/smpi_bench.cpp b/src/smpi/internals/smpi_bench.cpp index 85191c7329..af6cf7817b 100644 --- a/src/smpi/internals/smpi_bench.cpp +++ b/src/smpi/internals/smpi_bench.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2007-2018. 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. */ @@ -11,7 +11,6 @@ #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 "src/smpi/include/smpi_actor.hpp" @@ -28,6 +27,11 @@ 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; @@ -35,6 +39,7 @@ 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); @@ -51,7 +56,7 @@ void smpi_execute_flops(double 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()); @@ -96,12 +101,10 @@ void smpi_bench_begin() 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 @@ -177,7 +180,7 @@ 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(); @@ -232,6 +235,8 @@ int smpi_gettimeofday(struct timeval* tv, struct timezone* 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; } @@ -248,11 +253,28 @@ 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 () { @@ -314,9 +336,10 @@ 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 @@ -349,12 +372,11 @@ 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!"); @@ -362,21 +384,28 @@ 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) @@ -403,10 +432,7 @@ void smpi_sample_3(int global, const char *file, int line) 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, period); @@ -414,6 +440,25 @@ 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!"); + 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();