X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/dc1369e2055686dde3d178efb2001dbad8972d54..c49f2131b81992d4f990d9acb62dd8b365804fdc:/src/smpi/internals/smpi_global.cpp diff --git a/src/smpi/internals/smpi_global.cpp b/src/smpi/internals/smpi_global.cpp index 5f0e9887db..719aac55d4 100644 --- a/src/smpi/internals/smpi_global.cpp +++ b/src/smpi/internals/smpi_global.cpp @@ -1,50 +1,48 @@ -/* 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. */ #include "mc/mc.h" -#include "simgrid/s4u/Engine.hpp" +#include "simgrid/Exception.hpp" #include "simgrid/plugins/file_system.h" +#include "simgrid/s4u/Engine.hpp" #include "smpi_coll.hpp" +#include "smpi_config.hpp" #include "smpi_f2c.hpp" #include "smpi_host.hpp" +#include "src/kernel/EngineImpl.hpp" #include "src/kernel/activity/CommImpl.hpp" -#include "src/simix/smx_private.hpp" #include "src/smpi/include/smpi_actor.hpp" #include "xbt/config.hpp" +#include "xbt/file.hpp" #include -#include /* DBL_MAX */ +#include +#include /* split */ +#include +#include #include #include /* intmax_t */ +#include /* strerror */ #include #include #include +#include -#if not defined(__APPLE__) -#include +#if SG_HAVE_SENDFILE +#include #endif -#if defined(__APPLE__) -# include -# ifndef MAC_OS_X_VERSION_10_12 -# define MAC_OS_X_VERSION_10_12 101200 -# endif -constexpr bool HAVE_WORKING_MMAP = (MAC_OS_X_VERSION_MIN_REQUIRED >= MAC_OS_X_VERSION_10_12); -#elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) -constexpr bool HAVE_WORKING_MMAP = false; -#else -constexpr bool HAVE_WORKING_MMAP = true; +#if HAVE_PAPI +#include "papi.h" #endif -#if SG_HAVE_SENDFILE -#include +#if not defined(__APPLE__) && not defined(__HAIKU__) +#include #endif XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)"); -#include -#include /* trim_right / trim_left */ #if SMPI_IFORT extern "C" void for_rtl_init_ (int *, char **); @@ -67,48 +65,35 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (ke #endif #if HAVE_PAPI -#include "papi.h" -std::string papi_default_config_name = "default"; -std::map units2papi_setup; +std::map> units2papi_setup; #endif 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; -/** - * Setting MPI_COMM_WORLD to MPI_COMM_UNINITIALIZED (it's a variable) - * is important because the implementation of MPI_Comm checks - * "this == MPI_COMM_UNINITIALIZED"? If yes, it uses smpi_process()->comm_world() - * instead of "this". - * This is basically how we only have one global variable but all processes have - * different communicators (the one their SMPI instance uses). - * - * 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"); -static simgrid::config::Flag smpi_init_sleep( - "smpi/init", "Time to inject inside a call to MPI_Init", 0.0); + +static simgrid::config::Flag + smpi_hostfile("smpi/hostfile", + "Classical MPI hostfile containing list of machines to dispatch " + "the processes, one per line", + ""); + +static simgrid::config::Flag smpi_replay("smpi/replay", + "Replay a trace instead of executing the application", ""); + +static simgrid::config::Flag smpi_np("smpi/np", "Number of processes to be created", 0); + +static simgrid::config::Flag smpi_map("smpi/map", "Display the mapping between nodes and processes", 0); 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(); @@ -116,12 +101,14 @@ simgrid::smpi::ActorExt* smpi_process() if (me == nullptr) // This happens sometimes (eg, when linking against NS3 because it pulls openMPI...) return nullptr; - return process_data.at(me.get()); + return me->extension(); } simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor) { - return process_data.at(actor.get()); + if (actor.get() == nullptr) + return nullptr; + return actor->extension(); } MPI_Comm smpi_process_comm_self(){ @@ -132,16 +119,12 @@ MPI_Info smpi_process_info_env(){ return smpi_process()->info_env(); } -void smpi_process_init(int *argc, char ***argv){ - simgrid::smpi::ActorExt::init(); -} - void * smpi_process_get_user_data(){ - return simgrid::s4u::Actor::self()->get_impl()->get_user_data(); + return simgrid::s4u::Actor::self()->get_data(); } void smpi_process_set_user_data(void *data){ - simgrid::s4u::Actor::self()->get_impl()->set_user_data(data); + simgrid::s4u::Actor::self()->set_data(data); } void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t)) @@ -149,77 +132,82 @@ void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, s static void (*saved_callback)(smx_activity_t, void*, size_t); saved_callback = callback; smpi_comm_copy_data_callback = [](simgrid::kernel::activity::CommImpl* comm, void* buff, size_t size) { - saved_callback(smx_activity_t(comm), buff, size); + saved_callback(comm, buff, size); }; } -static void memcpy_private(void* dest, const void* src, std::vector>& private_blocks) +static void memcpy_private(void* dest, const void* src, const std::vector>& private_blocks) { for (auto const& block : private_blocks) memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first); } -static void check_blocks(std::vector> &private_blocks, size_t buff_size) { +static void check_blocks(const std::vector>& private_blocks, size_t buff_size) +{ for (auto const& block : private_blocks) xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc."); } +static void smpi_cleanup_comm_after_copy(simgrid::kernel::activity::CommImpl* comm, void* buff){ + if (comm->detached()) { + // if this is a detached send, the source buffer was duplicated by SMPI + // sender to make the original buffer available to the application ASAP + xbt_free(buff); + //It seems that the request is used after the call there this should be free somewhere else but where??? + //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free + comm->src_buff_ = nullptr; + } +} + void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff, size_t buff_size) { - int src_shared = 0; - int dst_shared = 0; size_t src_offset = 0; size_t dst_offset = 0; std::vector> src_private_blocks; std::vector> dst_private_blocks; XBT_DEBUG("Copy the data over"); - if((src_shared=smpi_is_shared(buff, src_private_blocks, &src_offset))) { - XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff); + if(smpi_is_shared(buff, src_private_blocks, &src_offset)) { src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size); + if (src_private_blocks.empty()) { // simple shared malloc ... return. + XBT_VERB("Sender is shared. Let's ignore it."); + smpi_cleanup_comm_after_copy(comm, buff); + return; + } } else { src_private_blocks.clear(); - src_private_blocks.push_back(std::make_pair(0, buff_size)); + src_private_blocks.emplace_back(0, buff_size); } - if ((dst_shared = smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset))) { - XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff_); + if (smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset)) { dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size); + if (dst_private_blocks.empty()) { // simple shared malloc ... return. + XBT_VERB("Receiver is shared. Let's ignore it."); + smpi_cleanup_comm_after_copy(comm, buff); + return; + } } else { dst_private_blocks.clear(); - dst_private_blocks.push_back(std::make_pair(0, buff_size)); + dst_private_blocks.emplace_back(0, buff_size); } check_blocks(src_private_blocks, buff_size); check_blocks(dst_private_blocks, buff_size); auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks); check_blocks(private_blocks, buff_size); void* tmpbuff=buff; - if ((smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) && - (static_cast(buff) >= smpi_data_exe_start) && - (static_cast(buff) < smpi_data_exe_start + smpi_data_exe_size)) { + if (smpi_switch_data_segment(comm->src_actor_->get_iface(), buff)) { XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !"); - smpi_switch_data_segment(comm->src_actor_->iface()); - tmpbuff = static_cast(xbt_malloc(buff_size)); + tmpbuff = xbt_malloc(buff_size); memcpy_private(tmpbuff, buff, private_blocks); } - if ((smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) && - ((char*)comm->dst_buff_ >= smpi_data_exe_start) && - ((char*)comm->dst_buff_ < smpi_data_exe_start + smpi_data_exe_size)) { + if (smpi_switch_data_segment(comm->dst_actor_->get_iface(), comm->dst_buff_)) XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment"); - smpi_switch_data_segment(comm->dst_actor_->iface()); - } + XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_); memcpy_private(comm->dst_buff_, tmpbuff, private_blocks); - if (comm->detached_) { - // if this is a detached send, the source buffer was duplicated by SMPI - // sender to make the original buffer available to the application ASAP - xbt_free(buff); - //It seems that the request is used after the call there this should be free somewhere else but where??? - //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free - comm->src_buff_ = nullptr; - } + smpi_cleanup_comm_after_copy(comm,buff); if (tmpbuff != buff) xbt_free(tmpbuff); } @@ -229,198 +217,86 @@ void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, v /* nothing done in this version */ } -static void smpi_check_options() -{ - //check correctness of MPI parameters - - xbt_assert(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")) { - 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." - "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information."); - } - - xbt_assert(simgrid::config::get_value("smpi/cpu-threshold") >= 0, - "The 'smpi/cpu-threshold' option cannot have negative values [anymore]. If you want to discard " - "the simulation of any computation, please use 'smpi/simulate-computation:no' instead."); -} - 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)) // and the (computed) event_set. - if (not simgrid::config::get_value("smpi/papi-events").empty()) { - if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT) - XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?" - " Expected version is %u", PAPI_VER_CURRENT); - - typedef boost::tokenizer> Tokenizer; - boost::char_separator separator_units(";"); - std::string str = simgrid::config::get_value("smpi/papi-events"); - Tokenizer tokens(str, separator_units); - - // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it. - // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units". - for (auto const& unit_it : tokens) { - boost::char_separator separator_events(":"); - Tokenizer event_tokens(unit_it, separator_events); - - int event_set = PAPI_NULL; - if (PAPI_create_eventset(&event_set) != PAPI_OK) { - // TODO: Should this let the whole simulation die? - XBT_CRITICAL("Could not create PAPI event set during init."); - } - - // NOTE: We cannot use a map here, as we must obey the order of the counters - // This is important for PAPI: We need to map the values of counters back - // to the event_names (so, when PAPI_read() has finished)! - papi_counter_t counters2values; - - // Iterate over all counters that were specified for this specific - // unit. - // Note that we need to remove the name of the unit - // (that could also be the "default" value), which always comes first. - // Hence, we start at ++(events.begin())! - for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) { - - int event_code = PAPI_NULL; - char* event_name = const_cast((*events_it).c_str()); - if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) { - XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name); - continue; - } - if (PAPI_add_event(event_set, event_code) != PAPI_OK) { - XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name); - continue; - } - XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name); - - counters2values.push_back( - // We cannot just pass *events_it, as this is of type const basic_string - std::make_pair(std::string(*events_it), 0)); - } - - std::string unit_name = *(event_tokens.begin()); - papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set}; - - units2papi_setup.insert(std::make_pair(unit_name, std::move(config))); - } - } -#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 (smpi_cfg_papi_events_file().empty()) + return; - if (not MC_is_active()) { - xbt_os_timer_free(global_timer); + if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT) { + XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked? Expected version is %u", + PAPI_VER_CURRENT); + return; } - if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) - smpi_destroy_global_memory_segments(); - if(simgrid::smpi::F2C::lookup() != nullptr) - simgrid::smpi::F2C::delete_lookup(); -} + using Tokenizer = boost::tokenizer>; + boost::char_separator separator_units(";"); + std::string str = smpi_cfg_papi_events_file(); + Tokenizer tokens(str, separator_units); + + // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it. + // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units". + for (auto const& unit_it : tokens) { + boost::char_separator separator_events(":"); + Tokenizer event_tokens(unit_it, separator_events); + + int event_set = PAPI_NULL; + if (PAPI_create_eventset(&event_set) != PAPI_OK) { + // TODO: Should this let the whole simulation die? + XBT_CRITICAL("Could not create PAPI event set during init."); + break; + } -static void smpi_init_options(){ - // return if already called - if (smpi_cpu_threshold > -1) - return; - 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); - 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; - else if (smpi_privatize_option == "yes" || smpi_privatize_option == "1") - smpi_privatize_global_variables = SmpiPrivStrategies::DEFAULT; - else if (smpi_privatize_option == "mmap") - smpi_privatize_global_variables = SmpiPrivStrategies::MMAP; - else if (smpi_privatize_option == "dlopen") - smpi_privatize_global_variables = SmpiPrivStrategies::DLOPEN; - else - xbt_die("Invalid value for smpi/privatization: '%s'", smpi_privatize_option.c_str()); + // NOTE: We cannot use a map here, as we must obey the order of the counters + // This is important for PAPI: We need to map the values of counters back to the event_names (so, when PAPI_read() + // has finished)! + papi_counter_t counters2values; + + // Iterate over all counters that were specified for this specific unit. + // Note that we need to remove the name of the unit (that could also be the "default" value), which always comes + // first. Hence, we start at ++(events.begin())! + for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) { + int event_code = PAPI_NULL; + auto* event_name = const_cast((*events_it).c_str()); + if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) { + XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name); + continue; + } + if (PAPI_add_event(event_set, event_code) != PAPI_OK) { + XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name); + continue; + } + XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name); - if (not SMPI_switch_data_segment) { - XBT_DEBUG("Running without smpi_main(); disable smpi/privatization."); - smpi_privatize_global_variables = SmpiPrivStrategies::NONE; - } - if (not HAVE_WORKING_MMAP && smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) { - XBT_INFO("mmap privatization is broken on this platform, switching to dlopen privatization instead."); - smpi_privatize_global_variables = SmpiPrivStrategies::DLOPEN; - } + counters2values.emplace_back(*events_it, 0LL); + } - if (smpi_cpu_threshold < 0) - smpi_cpu_threshold = DBL_MAX; + std::string unit_name = *(event_tokens.begin()); + papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set}; - 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; - } else if (val == "local") { - smpi_cfg_shared_malloc = SharedMallocType::LOCAL; - } else if ((val == "no") || (val == "0") || (val == "off")) { - smpi_cfg_shared_malloc = SharedMallocType::NONE; - } else { - xbt_die("Invalid value '%s' for option smpi/shared-malloc. Possible values: 'on' or 'global', 'local', 'off'", - val.c_str()); + units2papi_setup.insert(std::make_pair(unit_name, std::move(config))); } +#endif } -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)(); +using smpi_entry_point_type = std::function; +using smpi_c_entry_point_type = int (*)(int argc, char** argv); +using smpi_fortran_entry_point_type = void (*)(); -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, + const std::vector& args) { // copy C strings, we need them writable - std::vector* args4argv = new std::vector(args.size()); + auto* args4argv = new std::vector(args.size()); std::transform(begin(args), end(args), begin(*args4argv), [](const std::string& s) { return xbt_strdup(s.c_str()); }); // set argv[0] to executable_path @@ -431,11 +307,10 @@ static int smpi_run_entry_point(const smpi_entry_point_type& entry_point, const // take a copy of args4argv to keep reference of the allocated strings const std::vector args2str(*args4argv); #endif - int argc = args4argv->size(); + int argc = static_cast(args4argv->size()); args4argv->push_back(nullptr); char** argv = args4argv->data(); - simgrid::smpi::ActorExt::init(); #if SMPI_IFORT for_rtl_init_ (&argc, argv); #elif SMPI_FLANG @@ -443,8 +318,18 @@ static int smpi_run_entry_point(const smpi_entry_point_type& entry_point, const __io_set_argv(argv); #elif SMPI_GFORTRAN _gfortran_set_args(argc, argv); -#endif - int res = entry_point(argc, argv); +#endif + + try { + int res = entry_point(argc, argv); + if (res != 0) { + XBT_WARN("SMPI process did not return 0. Return value : %d", res); + if (smpi_exit_status == 0) + smpi_exit_status = res; + } + } catch (simgrid::ForcefulKillException const& e) { + XBT_DEBUG("Caught a ForcefulKillException: %s", e.what()); + } #if SMPI_IFORT for_rtl_finish_ (); @@ -454,11 +339,6 @@ static int smpi_run_entry_point(const smpi_entry_point_type& entry_point, const delete args4argv; #endif - if (res != 0){ - XBT_WARN("SMPI process did not return 0. Return value : %d", res); - if (smpi_exit_status == 0) - smpi_exit_status = res; - } return 0; } @@ -466,15 +346,15 @@ static int smpi_run_entry_point(const smpi_entry_point_type& entry_point, const // TODO, remove the number of functions involved here static smpi_entry_point_type smpi_resolve_function(void* handle) { - smpi_fortran_entry_point_type entry_point_fortran = (smpi_fortran_entry_point_type)dlsym(handle, "user_main_"); + auto* entry_point_fortran = reinterpret_cast(dlsym(handle, "user_main_")); if (entry_point_fortran != nullptr) { - return [entry_point_fortran](int argc, char** argv) { + return [entry_point_fortran](int, char**) { entry_point_fortran(); return 0; }; } - smpi_c_entry_point_type entry_point = (smpi_c_entry_point_type)dlsym(handle, "main"); + auto* entry_point = reinterpret_cast(dlsym(handle, "main")); if (entry_point != nullptr) { return entry_point; } @@ -486,54 +366,54 @@ static void smpi_copy_file(const std::string& src, const std::string& target, of { int fdin = open(src.c_str(), O_RDONLY); xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str()); - int fdout = open(target.c_str(), O_CREAT | O_RDWR, S_IRWXU); - xbt_assert(fdout >= 0, "Cannot write into %s", target.c_str()); + xbt_assert(unlink(target.c_str()) == 0 || errno == ENOENT, "Failed to unlink file %s: %s", target.c_str(), + strerror(errno)); + int fdout = open(target.c_str(), O_CREAT | O_RDWR | O_EXCL, S_IRWXU); + xbt_assert(fdout >= 0, "Cannot write into %s: %s", target.c_str(), strerror(errno)); XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast(fdin_size), target.c_str()); #if SG_HAVE_SENDFILE - ssize_t sent_size = sendfile(fdout, fdin, NULL, fdin_size); + ssize_t sent_size = sendfile(fdout, fdin, nullptr, fdin_size); if (sent_size == fdin_size) { close(fdin); close(fdout); return; - } else if (sent_size != -1 || errno != ENOSYS) { - xbt_die("Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(), - sent_size, static_cast(fdin_size), errno, strerror(errno)); } + xbt_assert(sent_size == -1 && errno == ENOSYS, + "Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(), + sent_size, static_cast(fdin_size), errno, strerror(errno)); #endif // If this point is reached, sendfile() actually is not available. Copy file by hand. - const int bufsize = 1024 * 1024 * 4; - char buf[bufsize]; - while (int got = read(fdin, buf, bufsize)) { + std::vector buf(1024 * 1024 * 4); + while (ssize_t got = read(fdin, buf.data(), buf.size())) { if (got == -1) { xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str()); - } else { - char* p = buf; - int todo = got; - while (int done = write(fdout, p, todo)) { - if (done == -1) { - xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str()); - } else { - p += done; - todo -= done; - } + continue; + } + const unsigned char* p = buf.data(); + ssize_t todo = got; + while (ssize_t done = write(fdout, p, todo)) { + if (done == -1) { + xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str()); + continue; } + p += done; + todo -= done; } } close(fdin); close(fdout); } -#if not defined(__APPLE__) +#if not defined(__APPLE__) && not defined(__HAIKU__) static int visit_libs(struct dl_phdr_info* info, size_t, void* data) { - char* libname = (char*)(data); - const char *path = info->dlpi_name; - if(strstr(path, libname)){ - strncpy(libname, path, 512); + auto* libname = static_cast(data); + std::string path = info->dlpi_name; + if (path.find(*libname) != std::string::npos) { + *libname = std::move(path); return 1; } - return 0; } #endif @@ -554,29 +434,29 @@ static void smpi_init_privatization_dlopen(const std::string& executable) for (auto const& libname : privatize_libs) { // load the library once to add it to the local libs, to get the absolute path void* libhandle = dlopen(libname.c_str(), RTLD_LAZY); + xbt_assert(libhandle != nullptr, "Cannot dlopen %s - check your settings in smpi/privatize-libs", + libname.c_str()); // get library name from path - char fullpath[512] = {'\0'}; - strncpy(fullpath, libname.c_str(), 511); -#if not defined(__APPLE__) - 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); + std::string fullpath = libname; +#if not defined(__APPLE__) && not defined(__HAIKU__) + xbt_assert(dl_iterate_phdr(visit_libs, &fullpath) != 0, + "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath.c_str()); + XBT_DEBUG("Extra lib to privatize '%s' found", fullpath.c_str()); #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); + privatize_libs_paths.emplace_back(std::move(fullpath)); dlclose(libhandle); } } - simix_global->default_function = [executable, fdin_size](std::vector args) { - return std::function([executable, fdin_size, args] { + simgrid::s4u::Engine::get_instance()->register_default([executable, fdin_size](std::vector args) { + return simgrid::kernel::actor::ActorCode([executable, fdin_size, args = std::move(args)] { static std::size_t rank = 0; // Copy the dynamic library: - std::string target_executable = - executable + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so"; + simgrid::xbt::Path path(executable); + std::string target_executable = simgrid::config::get_value("smpi/tmpdir") + "/" + + path.get_base_name() + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so"; smpi_copy_file(executable, target_executable, fdin_size); // if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one. @@ -596,19 +476,16 @@ static void smpi_init_privatization_dlopen(const std::string& executable) // Copy the dynamic library, the new name must be the same length as the old one // just replace the name with 7 digits for the rank and the rest of the name. - unsigned int pad = 7; - if (libname.length() < pad) - pad = libname.length(); - std::string target_lib = - std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad); + auto pad = std::min(7, libname.length()); + std::string target_libname = std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad); + std::string target_lib = simgrid::config::get_value("smpi/tmpdir") + "/" + target_libname; target_libs.push_back(target_lib); XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2); 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()); + std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_libname + "/g' " + target_executable; + int status = system(sedcommand.c_str()); + xbt_assert(status == 0, "error while applying sed command %s \n", sedcommand.c_str()); } } @@ -616,40 +493,51 @@ static void smpi_init_privatization_dlopen(const std::string& executable) // Load the copy and resolve the entry point: void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND); int saved_errno = errno; - if (simgrid::config::get_value("smpi/keep-temps") == false) { + if (not simgrid::config::get_value("smpi/keep-temps")) { unlink(target_executable.c_str()); 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).\nError: Did you compile the program with a SMPI-specific " + "compiler (spmicc or friends)?", + 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. Does your program contain a main() function?"); smpi_run_entry_point(entry_point, executable, args); }); - }; + }); } static void smpi_init_privatization_no_dlopen(const std::string& executable) { - if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) + if (smpi_cfg_privatization() == 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()); - if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) + xbt_assert(entry_point, "main not found in %s", executable.c_str()); + + if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) smpi_backup_global_memory_segment(); // Execute the same entry point for each simulated process: - simix_global->default_function = [entry_point, executable](std::vector args) { - return std::function( - [entry_point, executable, args] { smpi_run_entry_point(entry_point, executable, args); }); - }; + simgrid::s4u::Engine::get_instance()->register_default([entry_point, executable](std::vector args) { + return simgrid::kernel::actor::ActorCode([entry_point, executable, args = std::move(args)] { + if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) { + simgrid::smpi::ActorExt* ext = smpi_process(); + /* Now using the segment index of this process */ + ext->set_privatized_region(smpi_init_global_memory_segment_process()); + /* Done at the process's creation */ + smpi_switch_data_segment(simgrid::s4u::Actor::self()); + } + smpi_run_entry_point(entry_point, executable, args); + }); + }); } int smpi_main(const char* executable, int argc, char* argv[]) @@ -660,28 +548,30 @@ int smpi_main(const char* executable, int argc, char* argv[]) return 0; } - TRACE_global_init(); - SIMIX_global_init(&argc, argv); + smpi_init_options_internal(true); + simgrid::s4u::Engine engine(&argc, argv); - SMPI_switch_data_segment = &smpi_switch_data_segment; sg_storage_file_system_init(); // parse the platform file: get the host list - simgrid::s4u::Engine::get_instance()->load_platform(argv[1]); - SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback); + engine.load_platform(argv[1]); + engine.set_default_comm_data_copy_callback(smpi_comm_copy_buffer_callback); - smpi_init_options(); - if (smpi_privatize_global_variables == SmpiPrivStrategies::DLOPEN) + if (smpi_cfg_privatization() == SmpiPrivStrategies::DLOPEN) smpi_init_privatization_dlopen(executable); else smpi_init_privatization_no_dlopen(executable); + simgrid::smpi::colls::set_collectives(); + simgrid::smpi::colls::smpi_coll_cleanup_callback = nullptr; + SMPI_init(); - simgrid::s4u::Engine::get_instance()->load_deployment(argv[2]); - 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; + const std::vector args(argv + 2, argv + argc); + int rank_counts = + smpi_deployment_smpirun(&engine, smpi_hostfile.get(), smpi_np.get(), smpi_replay.get(), smpi_map.get(), args); + + SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, rank_counts); + MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name); /* Clean IO before the run */ fflush(stdout); @@ -690,55 +580,73 @@ int smpi_main(const char* executable, int argc, char* argv[]) if (MC_is_active()) { MC_run(); } else { - - SIMIX_run(); + engine.get_impl()->run(); xbt_os_walltimer_stop(global_timer); - if (simgrid::config::get_value("smpi/display-timing")) { - double global_time = xbt_os_timer_elapsed(global_timer); - XBT_INFO("Simulated time: %g seconds. \n\n" - "The simulation took %g seconds (after parsing and platform setup)\n" - "%g seconds were actual computation of the application", - SIMIX_get_clock(), global_time , smpi_total_benched_time); - - if (smpi_total_benched_time/global_time>=0.75) - XBT_INFO("More than 75%% of the time was spent inside the application code.\n" - "You may want to use sampling functions or trace replay to reduce this."); - } + simgrid::smpi::utils::print_time_analysis(xbt_os_timer_elapsed(global_timer)); } - smpi_global_destroy(); + SMPI_finalize(); return smpi_exit_status; } // Called either directly from the user code, or from the code called by smpirun void SMPI_init(){ + smpi_init_options_internal(false); 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)}); - } - }); - simgrid::s4u::Actor::on_destruction.connect([](simgrid::s4u::Actor const& actor) { - auto it = process_data.find(&actor); - if (it != process_data.end()) { - delete it->second; - process_data.erase(it); - } + if (not actor.is_daemon()) + actor.extension_set(new simgrid::smpi::ActorExt(&actor)); }); simgrid::s4u::Host::on_creation.connect( [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); }); + for (auto const& host : simgrid::s4u::Engine::get_instance()->get_all_hosts()) + 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); + } + 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_cfg_privatization() == SmpiPrivStrategies::MMAP) + smpi_destroy_global_memory_segments(); + + simgrid::smpi::utils::print_memory_analysis(); } void smpi_mpi_init() { smpi_init_fortran_types(); - if(smpi_init_sleep > 0) - simcall_process_sleep(smpi_init_sleep); + if(_smpi_init_sleep > 0) + simgrid::s4u::this_actor::sleep_for(_smpi_init_sleep); +} + +void SMPI_thread_create() { + TRACE_smpi_init(simgrid::s4u::this_actor::get_pid(), __func__); + smpi_process()->mark_as_initialized(); +} + +void smpi_exit(int res){ + if(res != 0){ + XBT_WARN("SMPI process did not return 0. Return value : %d", res); + smpi_exit_status = res; + } + simgrid::s4u::this_actor::exit(); + THROW_IMPOSSIBLE; }