X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/c215dd07037da91b2ef809d788e0535ba9a43789..52f3898881606b361670261250f095b0680c76de:/src/smpi/internals/smpi_global.cpp diff --git a/src/smpi/internals/smpi_global.cpp b/src/smpi/internals/smpi_global.cpp index 17e62ba88b..ab1ebfb125 100644 --- a/src/smpi/internals/smpi_global.cpp +++ b/src/smpi/internals/smpi_global.cpp @@ -9,15 +9,15 @@ #include "smpi_coll.hpp" #include "smpi_f2c.hpp" #include "smpi_host.hpp" +#include "smpi_config.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 -#include /* trim_right / trim_left */ +#include /* split */ #include -#include /* DBL_MAX */ #include #include /* intmax_t */ #include @@ -37,18 +37,6 @@ #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__) || defined(__sun) || defined(__HAIKU__) -constexpr bool HAVE_WORKING_MMAP = false; -#else -constexpr bool HAVE_WORKING_MMAP = true; -#endif - XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)"); #if SMPI_IFORT @@ -78,7 +66,6 @@ std::map units2papi_ std::unordered_map location2speedup; -static std::map process_data; static int smpi_exit_status = 0; extern double smpi_total_benched_time; xbt_os_timer_t global_timer; @@ -94,9 +81,6 @@ 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"); @@ -113,12 +97,14 @@ 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(){ @@ -130,11 +116,11 @@ MPI_Info smpi_process_info_env(){ } 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)) @@ -157,6 +143,17 @@ static void check_blocks(std::vector> &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) { size_t src_offset = 0; @@ -165,16 +162,24 @@ void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, v std::vector> dst_private_blocks; XBT_DEBUG("Copy the data over"); if(smpi_is_shared(buff, src_private_blocks, &src_offset)) { - XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff); src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size); + if (src_private_blocks.size()==0){//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)); } if (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_); dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size); + if (dst_private_blocks.size()==0){//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(); @@ -185,7 +190,7 @@ void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, v 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) && + if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) && (static_cast(buff) >= smpi_data_exe_start) && (static_cast(buff) < smpi_data_exe_start + smpi_data_exe_size)) { XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !"); @@ -194,7 +199,7 @@ void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, v memcpy_private(tmpbuff, buff, private_blocks); } - if ((smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) && + if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) && ((char*)comm->dst_buff_ >= smpi_data_exe_start) && ((char*)comm->dst_buff_ < smpi_data_exe_start + smpi_data_exe_size)) { XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment"); @@ -203,14 +208,7 @@ void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, v 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); } @@ -220,70 +218,25 @@ 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") && 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. " - "Check " - "https://simgrid.org/doc/latest/Configuring_SimGrid.html#automatic-benchmarking-of-smpi-code 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 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); - xbt_assert(fstream.is_open(), "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 (not smpi_cfg_papi_events_file().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"); + 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. @@ -324,7 +277,7 @@ void smpi_global_init() 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::make_pair(std::string(*events_it), 0LL)); } std::string unit_name = *(event_tokens.begin()); @@ -336,71 +289,7 @@ 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) - return; - simgrid::smpi::Colls::set_collectives(); - simgrid::smpi::Colls::smpi_coll_cleanup_callback = nullptr; - smpi_cpu_threshold = simgrid::config::get_value("smpi/cpu-threshold"); - 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; - 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()); - - 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; - } - - 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()); - } -} typedef std::function smpi_entry_point_type; typedef int (* smpi_c_entry_point_type)(int argc, char **argv); @@ -619,7 +508,7 @@ static void smpi_init_privatization_dlopen(const std::string& executable) 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: @@ -629,7 +518,7 @@ static void smpi_init_privatization_no_dlopen(const std::string& executable) smpi_entry_point_type entry_point = smpi_resolve_function(handle); xbt_assert(entry_point, "main not found in %s", executable.c_str()); - if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) + if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) smpi_backup_global_memory_segment(); // Execute the same entry point for each simulated process: @@ -646,26 +535,38 @@ int smpi_main(const char* executable, int argc, char* argv[]) * configuration tools */ return 0; } - + + SMPI_switch_data_segment = &smpi_switch_data_segment; + smpi_init_options(); TRACE_global_init(); SIMIX_global_init(&argc, argv); - SMPI_switch_data_segment = &smpi_switch_data_segment; + auto engine = simgrid::s4u::Engine::get_instance(); + sg_storage_file_system_init(); // parse the platform file: get the host list - simgrid::s4u::Engine::get_instance()->load_platform(argv[1]); + engine->load_platform(argv[1]); SIMIX_comm_set_copy_data_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... + + /* This is a ... heavy way to count the MPI ranks */ + int rank_counts = 0; + simgrid::s4u::Actor::on_creation.connect([&rank_counts](simgrid::s4u::Actor& actor) { + if (not actor.is_daemon()) + rank_counts++; + }); + engine->load_deployment(argv[2]); + + 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 */ @@ -691,39 +592,54 @@ 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; } // Called either directly from the user code, or from the code called by smpirun void SMPI_init(){ + smpi_init_options(); simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::Actor& actor) { - if (not actor.is_daemon()) { - process_data.insert({&actor, new simgrid::smpi::ActorExt(&actor)}); - } - }); - 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(); + 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); }