X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/e05a1337ff6756490bf19378bcaf875b002b9b82..749851ef42abca0a569ff59e8a8884fbac58f862:/src/smpi/internals/smpi_global.cpp diff --git a/src/smpi/internals/smpi_global.cpp b/src/smpi/internals/smpi_global.cpp index 47fa986abe..6ea962541c 100644 --- a/src/smpi/internals/smpi_global.cpp +++ b/src/smpi/internals/smpi_global.cpp @@ -3,27 +3,26 @@ /* 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/Mailbox.hpp" -#include "simgrid/s4u/Host.hpp" -#include "src/msg/msg_private.h" -#include "src/simix/smx_private.h" -#include "src/surf/surf_interface.hpp" #include "SmpiHost.hpp" -#include "xbt/config.hpp" -#include "private.h" +#include "mc/mc.h" #include "private.hpp" +#include "simgrid/s4u/Host.hpp" +#include "simgrid/s4u/Mailbox.hpp" #include "smpi_coll.hpp" #include "smpi_comm.hpp" #include "smpi_group.hpp" #include "smpi_info.hpp" #include "smpi_process.hpp" +#include "src/msg/msg_private.hpp" +#include "src/simix/smx_private.hpp" +#include "src/surf/surf_interface.hpp" +#include "xbt/config.hpp" +#include /* DBL_MAX */ #include #include -#include -#include /* DBL_MAX */ #include +#include #if HAVE_SENDFILE #include @@ -59,10 +58,23 @@ int smpi_universe_size = 0; int* index_to_process_data = nullptr; extern double smpi_total_benched_time; xbt_os_timer_t global_timer; +/** + * 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 (basically, 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 char* smpi_default_instance_name = "smpirun"; static simgrid::config::Flag smpi_wtime_sleep( "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0); static simgrid::config::Flag smpi_init_sleep( @@ -125,22 +137,20 @@ void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, s static void print(std::vector> vec) { std::fprintf(stderr, "{"); - for (auto elt : vec) { + for (auto const& elt : vec) { std::fprintf(stderr, "(0x%zx, 0x%zx),", elt.first, elt.second); - } - std::fprintf(stderr, "}\n"); + } + std::fprintf(stderr, "}\n"); } static void memcpy_private(void* dest, const void* src, std::vector>& private_blocks) { - for(auto block : 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) { - for(auto block : private_blocks) { + for (auto const& block : private_blocks) xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc."); - } } void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size) @@ -175,24 +185,23 @@ void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t b 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 == SMPI_PRIVATIZE_MMAP) && (static_cast(buff) >= smpi_start_data_exe) - && (static_cast(buff) < smpi_start_data_exe + smpi_size_data_exe ) - ){ - XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !"); - - smpi_switch_data_segment( - static_cast((static_cast(comm->src_proc->userdata)->data)) - ->index()); - tmpbuff = static_cast(xbt_malloc(buff_size)); - memcpy_private(tmpbuff, buff, private_blocks); + if ((smpi_privatize_global_variables == SMPI_PRIVATIZE_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 !"); + + smpi_switch_data_segment( + static_cast((static_cast(comm->src_proc->userdata)->data)) + ->index()); + tmpbuff = static_cast(xbt_malloc(buff_size)); + memcpy_private(tmpbuff, buff, private_blocks); } - if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_start_data_exe) - && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){ - XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment"); - smpi_switch_data_segment( - static_cast((static_cast(comm->dst_proc->userdata)->data)) - ->index()); + if ((smpi_privatize_global_variables == SMPI_PRIVATIZE_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"); + smpi_switch_data_segment( + static_cast((static_cast(comm->dst_proc->userdata)->data)) + ->index()); } XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff); memcpy_private(comm->dst_buff, tmpbuff, private_blocks); @@ -205,8 +214,8 @@ void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t b //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free comm->src_buff = nullptr; } - if(tmpbuff!=buff)xbt_free(tmpbuff); - + if (tmpbuff != buff) + xbt_free(tmpbuff); } void smpi_comm_null_copy_buffer_callback(smx_activity_t comm, void *buff, size_t buff_size) @@ -237,15 +246,13 @@ int smpi_enabled() { void smpi_global_init() { - MPI_Group group; - if (not MC_is_active()) { global_timer = xbt_os_timer_new(); xbt_os_walltimer_start(global_timer); } - if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') { - std::string filename {xbt_cfg_get_string("smpi/comp-adjustment-file")}; + std::string filename = xbt_cfg_get_string("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()); @@ -271,7 +278,7 @@ void smpi_global_init() // and the (computed) event_set. std::map units2papi_setup; - if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') { + if (not xbt_cfg_get_string("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 %i", @@ -279,14 +286,12 @@ void smpi_global_init() typedef boost::tokenizer> Tokenizer; boost::char_separator separator_units(";"); - std::string str = std::string(xbt_cfg_get_string("smpi/papi-events")); + std::string str = xbt_cfg_get_string("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& unit_it : tokens) { + // 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); @@ -306,7 +311,7 @@ void smpi_global_init() // 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++) { + 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()); @@ -335,43 +340,38 @@ void smpi_global_init() } #endif - int smpirun = 0; - msg_bar_t finalization_barrier = nullptr; - if (process_count == 0){ - process_count = SIMIX_process_count(); - smpirun=1; - finalization_barrier = MSG_barrier_init(process_count); + if (index_to_process_data == nullptr) { + index_to_process_data = new int[SIMIX_process_count()]; + } + + bool smpirun = 0; + if (process_count == 0) { // The program has been dispatched but no other + // SMPI instances have been registered. We're using smpirun. + smpirun = true; + SMPI_app_instance_register(smpi_default_instance_name, nullptr, + SIMIX_process_count()); // 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; process_data = new simgrid::smpi::Process*[process_count]; for (int i = 0; i < process_count; i++) { - process_data[i] = new simgrid::smpi::Process(i, finalization_barrier); - } - //if the process was launched through smpirun script we generate a global mpi_comm_world - //if not, we let MPI_COMM_NULL, and the comm world will be private to each mpi instance - if (smpirun) { - group = new simgrid::smpi::Group(process_count); - MPI_COMM_WORLD = new simgrid::smpi::Comm(group, nullptr); - MPI_Attr_put(MPI_COMM_WORLD, MPI_UNIVERSE_SIZE, reinterpret_cast(process_count)); - - for (int i = 0; i < process_count; i++) - group->set_mapping(i, i); + if (smpirun) { + process_data[i] = new simgrid::smpi::Process(i, smpi_deployment_finalization_barrier(smpi_default_instance_name)); + smpi_deployment_register_process(smpi_default_instance_name, i, i); + } else { + // TODO We can pass a nullptr here because Process::set_data() assigns the + // barrier from the instance anyway. This is ugly and should be changed + process_data[i] = new simgrid::smpi::Process(i, nullptr); + } } } void smpi_global_destroy() { - int count = smpi_process_count(); - smpi_bench_destroy(); smpi_shared_destroy(); - if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){ - delete MPI_COMM_WORLD->group(); - MSG_barrier_destroy(process_data[0]->finalization_barrier()); - }else{ - smpi_deployment_cleanup_instances(); - } - for (int i = 0; i < count; i++) { + smpi_deployment_cleanup_instances(); + for (int i = 0, count = smpi_process_count(); i < count; i++) { if(process_data[i]->comm_self()!=MPI_COMM_NULL){ simgrid::smpi::Comm::destroy(process_data[i]->comm_self()); } @@ -385,13 +385,8 @@ void smpi_global_destroy() delete[] process_data; process_data = nullptr; - if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){ - MPI_COMM_WORLD->cleanup_smp(); - MPI_COMM_WORLD->cleanup_attr(); - if(simgrid::smpi::Colls::smpi_coll_cleanup_callback!=nullptr) - simgrid::smpi::Colls::smpi_coll_cleanup_callback(); - delete MPI_COMM_WORLD; - } + if (simgrid::smpi::Colls::smpi_coll_cleanup_callback != nullptr) + simgrid::smpi::Colls::smpi_coll_cleanup_callback(); MPI_COMM_WORLD = MPI_COMM_NULL; @@ -399,7 +394,7 @@ void smpi_global_destroy() xbt_os_timer_free(global_timer); } - xbt_free(index_to_process_data); + delete[] index_to_process_data; if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) smpi_destroy_global_memory_segments(); smpi_free_static(); @@ -437,31 +432,24 @@ static void smpi_init_logs(){ } 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 = xbt_cfg_get_double("smpi/cpu-threshold"); - smpi_host_speed = xbt_cfg_get_double("smpi/host-speed"); - const char* smpi_privatize_option = xbt_cfg_get_string("smpi/privatization"); - if (std::strcmp(smpi_privatize_option, "no") == 0) - smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE; - else if (std::strcmp(smpi_privatize_option, "yes") == 0) - smpi_privatize_global_variables = SMPI_PRIVATIZE_DEFAULT; - else if (std::strcmp(smpi_privatize_option, "mmap") == 0) - smpi_privatize_global_variables = SMPI_PRIVATIZE_MMAP; - else if (std::strcmp(smpi_privatize_option, "dlopen") == 0) - smpi_privatize_global_variables = SMPI_PRIVATIZE_DLOPEN; - - // Some compatibility stuff: - else if (std::strcmp(smpi_privatize_option, "1") == 0) - smpi_privatize_global_variables = SMPI_PRIVATIZE_DEFAULT; - else if (std::strcmp(smpi_privatize_option, "0") == 0) - smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE; - - else - xbt_die("Invalid value for smpi/privatization: '%s'", smpi_privatize_option); + // 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 = xbt_cfg_get_double("smpi/cpu-threshold"); + smpi_host_speed = xbt_cfg_get_double("smpi/host-speed"); + std::string smpi_privatize_option = xbt_cfg_get_string("smpi/privatization"); + if (smpi_privatize_option == "no" || smpi_privatize_option == "0") + smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE; + else if (smpi_privatize_option == "yes" || smpi_privatize_option == "1") + smpi_privatize_global_variables = SMPI_PRIVATIZE_DEFAULT; + else if (smpi_privatize_option == "mmap") + smpi_privatize_global_variables = SMPI_PRIVATIZE_MMAP; + else if (smpi_privatize_option == "dlopen") + smpi_privatize_global_variables = SMPI_PRIVATIZE_DLOPEN; + else + xbt_die("Invalid value for smpi/privatization: '%s'", smpi_privatize_option.c_str()); #if defined(__FreeBSD__) if (smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) { @@ -473,7 +461,7 @@ static void smpi_init_options(){ if (smpi_cpu_threshold < 0) smpi_cpu_threshold = DBL_MAX; - char* val = xbt_cfg_get_string("smpi/shared-malloc"); + const char* val = xbt_cfg_get_string("smpi/shared-malloc").c_str(); if (not strcasecmp(val, "yes") || not strcmp(val, "1") || not strcasecmp(val, "on") || not strcasecmp(val, "global")) { smpi_cfg_shared_malloc = shmalloc_global; @@ -609,10 +597,11 @@ int smpi_main(const char* executable, int argc, char *argv[]) // Load the copy and resolve the entry point: void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND); + int saved_errno = errno; if (xbt_cfg_get_boolean("smpi/keep-temps") == false) unlink(target_executable.c_str()); if (handle == nullptr) - xbt_die("dlopen failed: %s (errno: %d -- %s)", dlerror(), errno, strerror(errno)); + xbt_die("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"); @@ -669,9 +658,8 @@ int smpi_main(const char* executable, int argc, char *argv[]) "You may want to use sampling functions or trace replay to reduce this."); } } - int count = smpi_process_count(); int ret = 0; - for (int i = 0; i < count; i++) { + for (int i = 0, count = smpi_process_count(); i < count; i++) { if(process_data[i]->return_value()!=0){ ret=process_data[i]->return_value();//return first non 0 value break; @@ -693,7 +681,7 @@ void SMPI_init(){ TRACE_smpi_alloc(); simgrid::surf::surfExitCallbacks.connect(TRACE_smpi_release); if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) - smpi_initialize_global_memory_segments(); + smpi_backup_global_memory_segment(); } void SMPI_finalize(){