X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/647b236abe6aacb7f3881369d4cb63e27b75aa8f..98755faee042e94d1ff52f6e9508b18015bb1ae5:/src/smpi/internals/smpi_global.cpp diff --git a/src/smpi/internals/smpi_global.cpp b/src/smpi/internals/smpi_global.cpp index c7ab73e5bc..8144ad9d14 100644 --- a/src/smpi/internals/smpi_global.cpp +++ b/src/smpi/internals/smpi_global.cpp @@ -25,6 +25,10 @@ #include #include +#if SIMGRID_HAVE_MC +#include "src/mc/mc_config.hpp" +#endif + #if SG_HAVE_SENDFILE #include #endif @@ -79,9 +83,7 @@ std::map units2papi_ std::unordered_map location2speedup; static std::map process_data; -int process_count = 0; static int smpi_exit_status = 0; -int smpi_universe_size = 0; extern double smpi_total_benched_time; xbt_os_timer_t global_timer; static std::vector privatize_libs_paths; @@ -96,9 +98,6 @@ static std::vector privatize_libs_paths; * See smpi_comm.cpp and the functions therein for details. */ MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED; -MPI_Errhandler *MPI_ERRORS_RETURN = nullptr; -MPI_Errhandler *MPI_ERRORS_ARE_FATAL = nullptr; -MPI_Errhandler *MPI_ERRHANDLER_NULL = nullptr; // No instance gets manually created; check also the smpirun.in script as // this default name is used there as well (when the tag is generated). static const std::string smpi_default_instance_name("smpirun"); @@ -108,11 +107,6 @@ static simgrid::config::Flag smpi_init_sleep( void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*, size_t) = &smpi_comm_copy_buffer_callback; -int smpi_process_count() -{ - return process_count; -} - simgrid::smpi::ActorExt* smpi_process() { simgrid::s4u::ActorPtr me = simgrid::s4u::Actor::self(); @@ -125,6 +119,8 @@ simgrid::smpi::ActorExt* smpi_process() simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor) { + if (actor.get() == nullptr) + return nullptr; return process_data.at(actor.get()); } @@ -136,11 +132,6 @@ MPI_Info smpi_process_info_env(){ return smpi_process()->info_env(); } -void smpi_process_init(int*, char***) -{ - simgrid::smpi::ActorExt::init(); -} - void * smpi_process_get_user_data(){ return simgrid::s4u::Actor::self()->get_impl()->get_user_data(); } @@ -234,12 +225,24 @@ void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, v static void smpi_check_options() { - //check correctness of MPI parameters +#if SIMGRID_HAVE_MC + if (MC_is_active()) { + if (_sg_mc_buffering == "zero") + simgrid::config::set_value("smpi/send-is-detached-thresh", 0); + else if (_sg_mc_buffering == "infty") + simgrid::config::set_value("smpi/send-is-detached-thresh", INT_MAX); + else + THROW_IMPOSSIBLE; + } +#endif xbt_assert(simgrid::config::get_value("smpi/async-small-thresh") <= + simgrid::config::get_value("smpi/send-is-detached-thresh"), + "smpi/async-small-thresh (=%d) should be smaller or equal to smpi/send-is-detached-thresh (=%d)", + simgrid::config::get_value("smpi/async-small-thresh"), simgrid::config::get_value("smpi/send-is-detached-thresh")); - if (simgrid::config::is_default("smpi/host-speed")) { + if (simgrid::config::is_default("smpi/host-speed") && not MC_is_active()) { XBT_INFO("You did not set the power of the host running the simulation. " "The timings will certainly not be accurate. " "Use the option \"--cfg=smpi/host-speed:\" to set its value. " @@ -254,37 +257,11 @@ static void smpi_check_options() } int smpi_enabled() { - return not process_data.empty(); + return MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED; } -void smpi_global_init() +static void smpi_init_papi() { - if (not MC_is_active()) { - global_timer = xbt_os_timer_new(); - xbt_os_walltimer_start(global_timer); - } - - std::string filename = simgrid::config::get_value("smpi/comp-adjustment-file"); - if (not filename.empty()) { - std::ifstream fstream(filename); - if (not fstream.is_open()) { - xbt_die("Could not open file %s. Does it exist?", filename.c_str()); - } - - std::string line; - typedef boost::tokenizer< boost::escaped_list_separator> Tokenizer; - std::getline(fstream, line); // Skip the header line - while (std::getline(fstream, line)) { - Tokenizer tok(line); - Tokenizer::iterator it = tok.begin(); - Tokenizer::iterator end = std::next(tok.begin()); - - std::string location = *it; - boost::trim(location); - location2speedup.insert(std::pair(location, std::stod(*end))); - } - } - #if HAVE_PAPI // This map holds for each computation unit (such as "default" or "process1" etc.) // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter)) @@ -350,27 +327,6 @@ void smpi_global_init() #endif } -void smpi_global_destroy() -{ - smpi_bench_destroy(); - smpi_shared_destroy(); - smpi_deployment_cleanup_instances(); - - if (simgrid::smpi::Colls::smpi_coll_cleanup_callback != nullptr) - simgrid::smpi::Colls::smpi_coll_cleanup_callback(); - - MPI_COMM_WORLD = MPI_COMM_NULL; - - if (not MC_is_active()) { - xbt_os_timer_free(global_timer); - } - - if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) - smpi_destroy_global_memory_segments(); - if(simgrid::smpi::F2C::lookup() != nullptr) - simgrid::smpi::F2C::delete_lookup(); -} - static void smpi_init_options(){ // return if already called if (smpi_cpu_threshold > -1) @@ -378,8 +334,10 @@ static void smpi_init_options(){ simgrid::smpi::Colls::set_collectives(); simgrid::smpi::Colls::smpi_coll_cleanup_callback = nullptr; smpi_cpu_threshold = simgrid::config::get_value("smpi/cpu-threshold"); - smpi_host_speed = simgrid::config::get_value("smpi/host-speed"); - xbt_assert(smpi_host_speed > 0.0, "You're trying to set the host_speed to a non-positive value (given: %f)", smpi_host_speed); + if (smpi_cpu_threshold < 0) + smpi_cpu_threshold = DBL_MAX; + + smpi_host_speed = simgrid::config::get_value("smpi/host-speed"); std::string smpi_privatize_option = simgrid::config::get_value("smpi/privatization"); if (smpi_privatize_option == "no" || smpi_privatize_option == "0") smpi_privatize_global_variables = SmpiPrivStrategies::NONE; @@ -401,9 +359,6 @@ static void smpi_init_options(){ smpi_privatize_global_variables = SmpiPrivStrategies::DLOPEN; } - if (smpi_cpu_threshold < 0) - smpi_cpu_threshold = DBL_MAX; - std::string val = simgrid::config::get_value("smpi/shared-malloc"); if ((val == "yes") || (val == "1") || (val == "on") || (val == "global")) { smpi_cfg_shared_malloc = SharedMallocType::GLOBAL; @@ -421,8 +376,8 @@ typedef std::function smpi_entry_point_type; typedef int (* smpi_c_entry_point_type)(int argc, char **argv); typedef void (*smpi_fortran_entry_point_type)(); -static int smpi_run_entry_point(const smpi_entry_point_type& entry_point, const std::string& executable_path, - std::vector args) +template +static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path, std::vector args) { // copy C strings, we need them writable std::vector* args4argv = new std::vector(args.size()); @@ -440,7 +395,6 @@ static int smpi_run_entry_point(const smpi_entry_point_type& entry_point, const args4argv->push_back(nullptr); char** argv = args4argv->data(); - simgrid::smpi::ActorExt::init(); #if SMPI_IFORT for_rtl_init_ (&argc, argv); #elif SMPI_FLANG @@ -564,13 +518,11 @@ static void smpi_init_privatization_dlopen(const std::string& executable) char fullpath[512] = {'\0'}; strncpy(fullpath, libname.c_str(), 511); #if not defined(__APPLE__) && not defined(__HAIKU__) - int ret = dl_iterate_phdr(visit_libs, fullpath); - if (ret == 0) - xbt_die("Can't find a linked %s - check the setting you gave to smpi/privatize-libs", fullpath); - else - XBT_DEBUG("Extra lib to privatize found : %s", fullpath); + xbt_assert(0 != dl_iterate_phdr(visit_libs, fullpath), + "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath); + XBT_DEBUG("Extra lib to privatize '%s' found", fullpath); #else - xbt_die("smpi/privatize-libs is not (yet) compatible with OSX"); + xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku"); #endif privatize_libs_paths.push_back(fullpath); dlclose(libhandle); @@ -612,9 +564,7 @@ static void smpi_init_privatization_dlopen(const std::string& executable) smpi_copy_file(libpath, target_lib, fdin_size2); std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_lib + "/g' " + target_executable; - int ret = system(sedcommand.c_str()); - if (ret != 0) - xbt_die("error while applying sed command %s \n", sedcommand.c_str()); + xbt_assert(system(sedcommand.c_str()) == 0, "error while applying sed command %s \n", sedcommand.c_str()); } } @@ -627,11 +577,11 @@ static void smpi_init_privatization_dlopen(const std::string& executable) for (const std::string& target_lib : target_libs) unlink(target_lib.c_str()); } - if (handle == nullptr) - xbt_die("dlopen failed: %s (errno: %d -- %s)", dlerror(), saved_errno, strerror(saved_errno)); + xbt_assert(handle != nullptr, "dlopen failed: %s (errno: %d -- %s)", dlerror(), saved_errno, + strerror(saved_errno)); + smpi_entry_point_type entry_point = smpi_resolve_function(handle); - if (not entry_point) - xbt_die("Could not resolve entry point"); + xbt_assert(entry_point, "Could not resolve entry point"); smpi_run_entry_point(entry_point, executable, args); }); }; @@ -641,13 +591,14 @@ static void smpi_init_privatization_no_dlopen(const std::string& executable) { if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) smpi_prepare_global_memory_segment(); + // Load the dynamic library and resolve the entry point: void* handle = dlopen(executable.c_str(), RTLD_LAZY | RTLD_LOCAL); - if (handle == nullptr) - xbt_die("dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno, strerror(errno)); + xbt_assert(handle != nullptr, "dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno, + strerror(errno)); smpi_entry_point_type entry_point = smpi_resolve_function(handle); - if (not entry_point) - xbt_die("main not found in %s", executable.c_str()); + xbt_assert(entry_point, "main not found in %s", executable.c_str()); + if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) smpi_backup_global_memory_segment(); @@ -686,8 +637,6 @@ int smpi_main(const char* executable, int argc, char* argv[]) SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, process_data.size()); // This call has a side effect on process_count... MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name); - smpi_universe_size = process_count; - /* Clean IO before the run */ fflush(stdout); @@ -712,7 +661,7 @@ int smpi_main(const char* executable, int argc, char* argv[]) "You may want to use sampling functions or trace replay to reduce this."); } } - smpi_global_destroy(); + SMPI_finalize(); return smpi_exit_status; } @@ -721,10 +670,11 @@ int smpi_main(const char* executable, int argc, char* argv[]) void SMPI_init(){ simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::Actor& actor) { if (not actor.is_daemon()) { - process_data.insert({&actor, new simgrid::smpi::ActorExt(&actor, nullptr)}); + process_data.insert({&actor, new simgrid::smpi::ActorExt(&actor)}); } }); simgrid::s4u::Actor::on_destruction.connect([](simgrid::s4u::Actor const& actor) { + XBT_DEBUG("Delete the extension of actor %s", actor.get_cname()); auto it = process_data.find(&actor); if (it != process_data.end()) { delete it->second; @@ -735,16 +685,56 @@ void SMPI_init(){ [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); }); smpi_init_options(); - smpi_global_init(); + if (not MC_is_active()) { + global_timer = xbt_os_timer_new(); + xbt_os_walltimer_start(global_timer); + } + + std::string filename = simgrid::config::get_value("smpi/comp-adjustment-file"); + if (not filename.empty()) { + std::ifstream fstream(filename); + xbt_assert(fstream.is_open(), "Could not open file %s. Does it exist?", filename.c_str()); + + std::string line; + typedef boost::tokenizer> Tokenizer; + std::getline(fstream, line); // Skip the header line + while (std::getline(fstream, line)) { + Tokenizer tok(line); + Tokenizer::iterator it = tok.begin(); + Tokenizer::iterator end = std::next(tok.begin()); + + std::string location = *it; + boost::trim(location); + location2speedup.insert(std::pair(location, std::stod(*end))); + } + } + smpi_init_papi(); smpi_check_options(); } -void SMPI_finalize(){ - smpi_global_destroy(); +void SMPI_finalize() +{ + smpi_bench_destroy(); + smpi_shared_destroy(); + smpi_deployment_cleanup_instances(); + + if (simgrid::smpi::Colls::smpi_coll_cleanup_callback != nullptr) + simgrid::smpi::Colls::smpi_coll_cleanup_callback(); + + MPI_COMM_WORLD = MPI_COMM_NULL; + + if (not MC_is_active()) { + xbt_os_timer_free(global_timer); + } + + if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) + smpi_destroy_global_memory_segments(); + if (simgrid::smpi::F2C::lookup() != nullptr) + simgrid::smpi::F2C::delete_lookup(); } void smpi_mpi_init() { smpi_init_fortran_types(); if(smpi_init_sleep > 0) - simcall_process_sleep(smpi_init_sleep); + simgrid::s4u::this_actor::sleep_for(smpi_init_sleep); }