X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/d70128f70514554704db4f73ce52cde5b8380b5e..e303bc683c874808621082e409ec13bcc1b17247:/src/smpi/smpi_global.cpp diff --git a/src/smpi/smpi_global.cpp b/src/smpi/smpi_global.cpp index 924e89b011..25518210f0 100644 --- a/src/smpi/smpi_global.cpp +++ b/src/smpi/smpi_global.cpp @@ -1,22 +1,30 @@ -/* Copyright (c) 2007-2015. The SimGrid Team. - * All rights reserved. */ +/* Copyright (c) 2007-2017. 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 +#include +#include +#include +#include +#include + #include "mc/mc.h" #include "private.h" #include "private.hpp" #include "simgrid/s4u/Mailbox.hpp" #include "simgrid/sg_config.h" -#include "smpi_mpi_dt_private.h" #include "src/kernel/activity/SynchroComm.hpp" #include "src/mc/mc_record.h" #include "src/mc/mc_replay.h" #include "src/msg/msg_private.h" #include "src/simix/smx_private.h" +#include "src/surf/surf_interface.hpp" +#include "src/smpi/SmpiHost.hpp" #include "surf/surf.h" -#include "xbt/replay.h" +#include "xbt/replay.hpp" +#include #include /* DBL_MAX */ #include @@ -25,12 +33,29 @@ #include #include #include +#include #include +#include + +#if HAVE_SENDFILE +#include +#endif XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)"); #include #include /* trim_right / trim_left */ +#ifndef RTLD_DEEPBIND +/* RTLD_DEEPBIND is a bad idea of GNU ld that obviously does not exist on other platforms + * See https://www.akkadia.org/drepper/dsohowto.pdf + * and https://lists.freebsd.org/pipermail/freebsd-current/2016-March/060284.html +*/ +#define RTLD_DEEPBIND 0 +#endif + +/* Mac OSX does not have any header file providing that definition so we have to duplicate it here. Bummers. */ +extern char** environ; /* we use it in posix_spawnp below */ + #if HAVE_PAPI #include "papi.h" const char* papi_default_config_name = "default"; @@ -43,180 +68,53 @@ struct papi_process_data { #endif std::unordered_map location2speedup; -typedef struct s_smpi_process_data { - double simulated; - int *argc; - char ***argv; - simgrid::s4u::MailboxPtr mailbox; - simgrid::s4u::MailboxPtr mailbox_small; - xbt_mutex_t mailboxes_mutex; - xbt_os_timer_t timer; - MPI_Comm comm_self; - MPI_Comm comm_intra; - MPI_Comm* comm_world; - void *data; /* user data */ - int index; - char state; - int sampling; /* inside an SMPI_SAMPLE_ block? */ - char* instance_id; - bool replaying; /* is the process replaying a trace */ - msg_bar_t finalization_barrier; - int return_value; - smpi_trace_call_location_t trace_call_loc; -#if HAVE_PAPI - /** Contains hardware data as read by PAPI **/ - int papi_event_set; - papi_counter_t papi_counter_data; -#endif -} s_smpi_process_data_t; - -static smpi_process_data_t *process_data = nullptr; +simgrid::smpi::Process **process_data = nullptr; int process_count = 0; int smpi_universe_size = 0; int* index_to_process_data = nullptr; extern double smpi_total_benched_time; -extern xbt_dict_t smpi_type_keyvals; -extern xbt_dict_t smpi_comm_keyvals; xbt_os_timer_t global_timer; 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; +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( + "smpi/init", "Time to inject inside a call to MPI_Init", 0.0); void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t) = &smpi_comm_copy_buffer_callback; -#define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1) - -static char *get_mailbox_name(char *str, int index) -{ - snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", static_cast (sizeof(int) * 2), index); - return str; -} - -static char *get_mailbox_name_small(char *str, int index) -{ - snprintf(str, MAILBOX_NAME_MAXLEN, "small%0*x", static_cast (sizeof(int) * 2), index); - return str; -} -void smpi_process_init(int *argc, char ***argv) -{ - if (process_data == nullptr){ - printf("SimGrid was not initialized properly before entering MPI_Init. Aborting, please check compilation process and use smpirun\n"); - exit(1); - } - if (argc != nullptr && argv != nullptr) { - smx_actor_t proc = SIMIX_process_self(); - proc->context->set_cleanup(&MSG_process_cleanup_from_SIMIX); - char* instance_id = (*argv)[1]; - int rank = xbt_str_parse_int((*argv)[2], "Invalid rank: %s"); - int index = smpi_process_index_of_smx_process(proc); - - if(index_to_process_data == nullptr){ - index_to_process_data=static_cast(xbt_malloc(SIMIX_process_count()*sizeof(int))); - } - - if(smpi_privatize_global_variables){ - /* Now using segment index of the process */ - index = proc->segment_index; - /* Done at the process's creation */ - SMPI_switch_data_segment(index); - } - - MPI_Comm* temp_comm_world; - msg_bar_t temp_bar; - smpi_deployment_register_process(instance_id, rank, index, &temp_comm_world, &temp_bar); - smpi_process_data_t data = smpi_process_remote_data(index); - data->comm_world = temp_comm_world; - if(temp_bar != nullptr) - data->finalization_barrier = temp_bar; - data->index = index; - data->instance_id = instance_id; - data->replaying = false; - - static_cast(proc->data)->data = data; - - if (*argc > 3) { - memmove(&(*argv)[0], &(*argv)[2], sizeof(char *) * (*argc - 2)); - (*argv)[(*argc) - 1] = nullptr; - (*argv)[(*argc) - 2] = nullptr; - } - (*argc)-=2; - data->argc = argc; - data->argv = argv; - // set the process attached to the mailbox - data->mailbox_small->setReceiver(simgrid::s4u::Actor::self()); - XBT_DEBUG("<%d> New process in the game: %p", index, proc); - } - xbt_assert(smpi_process_data(), - "smpi_process_data() returned nullptr. You probably gave a nullptr parameter to MPI_Init. " - "Although it's required by MPI-2, this is currently not supported by SMPI."); -} - -void smpi_process_destroy() +int smpi_process_count() { - int index = smpi_process_index(); - if(smpi_privatize_global_variables){ - smpi_switch_data_segment(index); - } - process_data[index_to_process_data[index]]->state = SMPI_FINALIZED; - XBT_DEBUG("<%d> Process left the game", index); + return process_count; } -/** @brief Prepares the current process for termination. */ -void smpi_process_finalize() +simgrid::smpi::Process* smpi_process() { - // This leads to an explosion of the search graph which cannot be reduced: - if(MC_is_active() || MC_record_replay_is_active()) - return; - - int index = smpi_process_index(); - // wait for all pending asynchronous comms to finish - MSG_barrier_wait(process_data[index_to_process_data[index]]->finalization_barrier); + simgrid::MsgActorExt* msgExt = static_cast(SIMIX_process_self()->data); + return static_cast(msgExt->data); } -/** @brief Check if a process is finalized */ -int smpi_process_finalized() +simgrid::smpi::Process* smpi_process_remote(int index) { - int index = smpi_process_index(); - if (index != MPI_UNDEFINED) - return (process_data[index_to_process_data[index]]->state == SMPI_FINALIZED); - else - return 0; + return process_data[index_to_process_data[index]]; } -/** @brief Check if a process is initialized */ -int smpi_process_initialized() -{ - if (index_to_process_data == nullptr){ - return false; - } else{ - int index = smpi_process_index(); - return ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state == SMPI_INITIALIZED)); - } +MPI_Comm smpi_process_comm_self(){ + return smpi_process()->comm_self(); } -/** @brief Mark a process as initialized (=MPI_Init called) */ -void smpi_process_mark_as_initialized() -{ - int index = smpi_process_index(); - if ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state != SMPI_FINALIZED)) - process_data[index_to_process_data[index]]->state = SMPI_INITIALIZED; +void smpi_process_init(int *argc, char ***argv){ + simgrid::smpi::Process::init(argc, argv); } -void smpi_process_set_replaying(bool value){ - int index = smpi_process_index(); - if ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state != SMPI_FINALIZED)) - process_data[index_to_process_data[index]]->replaying = value; +int smpi_process_index(){ + return smpi_process()->index(); } -bool smpi_process_get_replaying(){ - int index = smpi_process_index(); - if (index != MPI_UNDEFINED) - return process_data[index_to_process_data[index]]->replaying; - else return (_xbt_replay_is_active() != 0); -} int smpi_global_size() { @@ -226,170 +124,6 @@ int smpi_global_size() return xbt_str_parse_int(value, "SMPI_GLOBAL_SIZE contains a non-numerical value: %s"); } -smpi_process_data_t smpi_process_data() -{ - simgrid::MsgActorExt* msgExt = static_cast(SIMIX_process_self()->data); - return static_cast(msgExt->data); -} - -smpi_process_data_t smpi_process_remote_data(int index) -{ - return process_data[index_to_process_data[index]]; -} - -void smpi_process_set_user_data(void *data) -{ - smpi_process_data_t process_data = smpi_process_data(); - process_data->data = data; -} - -void *smpi_process_get_user_data() -{ - smpi_process_data_t process_data = smpi_process_data(); - return process_data->data; -} - -int smpi_process_count() -{ - return process_count; -} - -/** - * \brief Returns a structure that stores the location (filename + linenumber) - * of the last calls to MPI_* functions. - * - * \see smpi_trace_set_call_location - */ -smpi_trace_call_location_t* smpi_process_get_call_location() -{ - smpi_process_data_t process_data = smpi_process_data(); - return &process_data->trace_call_loc; -} - -int smpi_process_index() -{ - smpi_process_data_t data = smpi_process_data(); - //return -1 if not initialized - return data != nullptr ? data->index : MPI_UNDEFINED; -} - -MPI_Comm smpi_process_comm_world() -{ - smpi_process_data_t data = smpi_process_data(); - //return MPI_COMM_NULL if not initialized - return data != nullptr ? *data->comm_world : MPI_COMM_NULL; -} - -smx_mailbox_t smpi_process_mailbox() -{ - smpi_process_data_t data = smpi_process_data(); - return data->mailbox->getImpl(); -} - -smx_mailbox_t smpi_process_mailbox_small() -{ - smpi_process_data_t data = smpi_process_data(); - return data->mailbox_small->getImpl(); -} - -xbt_mutex_t smpi_process_mailboxes_mutex() -{ - smpi_process_data_t data = smpi_process_data(); - return data->mailboxes_mutex; -} - -smx_mailbox_t smpi_process_remote_mailbox(int index) -{ - smpi_process_data_t data = smpi_process_remote_data(index); - return data->mailbox->getImpl(); -} - -smx_mailbox_t smpi_process_remote_mailbox_small(int index) -{ - smpi_process_data_t data = smpi_process_remote_data(index); - return data->mailbox_small->getImpl(); -} - -xbt_mutex_t smpi_process_remote_mailboxes_mutex(int index) -{ - smpi_process_data_t data = smpi_process_remote_data(index); - return data->mailboxes_mutex; -} - -#if HAVE_PAPI -int smpi_process_papi_event_set(void) -{ - smpi_process_data_t data = smpi_process_data(); - return data->papi_event_set; -} - -papi_counter_t& smpi_process_papi_counters(void) -{ - smpi_process_data_t data = smpi_process_data(); - return data->papi_counter_data; -} -#endif - -xbt_os_timer_t smpi_process_timer() -{ - smpi_process_data_t data = smpi_process_data(); - return data->timer; -} - -void smpi_process_simulated_start() -{ - smpi_process_data_t data = smpi_process_data(); - data->simulated = SIMIX_get_clock(); -} - -double smpi_process_simulated_elapsed() -{ - smpi_process_data_t data = smpi_process_data(); - return SIMIX_get_clock() - data->simulated; -} - -MPI_Comm smpi_process_comm_self() -{ - smpi_process_data_t data = smpi_process_data(); - if(data->comm_self==MPI_COMM_NULL){ - MPI_Group group = new simgrid::SMPI::Group(1); - data->comm_self = new simgrid::SMPI::Comm(group, nullptr); - group->set_mapping(smpi_process_index(), 0); - } - - return data->comm_self; -} - -MPI_Comm smpi_process_get_comm_intra() -{ - smpi_process_data_t data = smpi_process_data(); - return data->comm_intra; -} - -void smpi_process_set_comm_intra(MPI_Comm comm) -{ - smpi_process_data_t data = smpi_process_data(); - data->comm_intra = comm; -} - -void smpi_process_set_sampling(int s) -{ - smpi_process_data_t data = smpi_process_data(); - data->sampling = s; -} - -int smpi_process_get_sampling() -{ - smpi_process_data_t data = smpi_process_data(); - return data->sampling; -} - -void print_request(const char *message, MPI_Request request) -{ - XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]", - message, request, request->buf, request->size, request->src, request->dst, request->tag, request->flags); -} - void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t)) { smpi_comm_copy_data_callback = callback; @@ -397,39 +131,48 @@ void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, s void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size) { - XBT_DEBUG("Copy the data over"); - void* tmpbuff=buff; + simgrid::kernel::activity::Comm *comm = dynamic_cast(synchro); - if((smpi_privatize_global_variables) && (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 !"); + XBT_DEBUG("Copy the data over"); + if(smpi_is_shared(buff)){ + XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff); + }else if(smpi_is_shared((char*)comm->dst_buff)){ + XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff); + }else{ + 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->data)->data))->index())); + tmpbuff = static_cast(xbt_malloc(buff_size)); + memcpy(tmpbuff, buff, buff_size); + } - smpi_switch_data_segment( - (static_cast((static_cast(comm->src_proc->data)->data))->index)); - tmpbuff = static_cast(xbt_malloc(buff_size)); - memcpy(tmpbuff, buff, buff_size); - } + 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->data)->data))->index())); + } - if((smpi_privatize_global_variables) && ((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->data)->data))->index)); - } + XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff); + memcpy(comm->dst_buff, tmpbuff, buff_size); - memcpy(comm->dst_buff, tmpbuff, buff_size); - 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; + 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; + } + 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) @@ -460,10 +203,7 @@ int smpi_enabled() { void smpi_global_init() { - int i; MPI_Group group; - char name[MAILBOX_NAME_MAXLEN]; - int smpirun=0; if (!MC_is_active()) { global_timer = xbt_os_timer_new(); @@ -560,59 +300,28 @@ 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); } smpi_universe_size = process_count; - process_data = new smpi_process_data_t[process_count]; - for (i = 0; i < process_count; i++) { - process_data[i] = new s_smpi_process_data_t; - process_data[i]->argc = nullptr; - process_data[i]->argv = nullptr; - process_data[i]->mailbox = simgrid::s4u::Mailbox::byName(get_mailbox_name(name, i)); - process_data[i]->mailbox_small = simgrid::s4u::Mailbox::byName(get_mailbox_name_small(name, i)); - process_data[i]->mailboxes_mutex = xbt_mutex_init(); - process_data[i]->timer = xbt_os_timer_new(); - if (MC_is_active()) - MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size()); - process_data[i]->comm_self = MPI_COMM_NULL; - process_data[i]->comm_intra = MPI_COMM_NULL; - process_data[i]->comm_world = nullptr; - process_data[i]->state = SMPI_UNINITIALIZED; - process_data[i]->sampling = 0; - process_data[i]->finalization_barrier = nullptr; - process_data[i]->return_value = 0; - -#if HAVE_PAPI - if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') { - // TODO: Implement host/process/thread based counters. This implementation - // just always takes the values passed via "default", like this: - // "default:COUNTER1:COUNTER2:COUNTER3;". - auto it = units2papi_setup.find(papi_default_config_name); - if (it != units2papi_setup.end()) { - process_data[i]->papi_event_set = it->second.event_set; - process_data[i]->papi_counter_data = it->second.counter_data; - XBT_DEBUG("Setting PAPI set for process %i", i); - } else { - process_data[i]->papi_event_set = PAPI_NULL; - XBT_DEBUG("No PAPI set for process %i", i); - } - } -#endif + 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); + 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)); - msg_bar_t bar = MSG_barrier_init(process_count); - for (i = 0; i < process_count; i++) { + for (int i = 0; i < process_count; i++) group->set_mapping(i, i); - process_data[i]->finalization_barrier = bar; - } } } @@ -621,21 +330,22 @@ void smpi_global_destroy() int count = smpi_process_count(); smpi_bench_destroy(); + smpi_shared_destroy(); if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){ - while (MPI_COMM_WORLD->group()->unuse() > 0); - MSG_barrier_destroy(process_data[0]->finalization_barrier); + 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++) { - if(process_data[i]->comm_self!=MPI_COMM_NULL){ - process_data[i]->comm_self->destroy(); + if(process_data[i]->comm_self()!=MPI_COMM_NULL){ + simgrid::smpi::Comm::destroy(process_data[i]->comm_self()); } - if(process_data[i]->comm_intra!=MPI_COMM_NULL){ - process_data[i]->comm_intra->destroy(); + if(process_data[i]->comm_intra()!=MPI_COMM_NULL){ + simgrid::smpi::Comm::destroy(process_data[i]->comm_intra()); } - xbt_os_timer_free(process_data[i]->timer); - xbt_mutex_destroy(process_data[i]->mailboxes_mutex); + xbt_os_timer_free(process_data[i]->timer()); + xbt_mutex_destroy(process_data[i]->mailboxes_mutex()); delete process_data[i]; } delete[] process_data; @@ -643,10 +353,10 @@ void smpi_global_destroy() if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){ MPI_COMM_WORLD->cleanup_smp(); - MPI_COMM_WORLD->cleanup_attributes(); - if(smpi_coll_cleanup_callback!=nullptr) - smpi_coll_cleanup_callback(); - xbt_free(MPI_COMM_WORLD); + 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; } MPI_COMM_WORLD = MPI_COMM_NULL; @@ -656,46 +366,13 @@ void smpi_global_destroy() } xbt_free(index_to_process_data); - if(smpi_type_keyvals!=nullptr) - xbt_dict_free(&smpi_type_keyvals); - if(smpi_comm_keyvals!=nullptr) - xbt_dict_free(&smpi_comm_keyvals); - if(smpi_privatize_global_variables) + if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) smpi_destroy_global_memory_segments(); smpi_free_static(); } -#ifndef WIN32 - -void __attribute__ ((weak)) user_main_() -{ - xbt_die("Should not be in this smpi_simulated_main"); -} - -int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv) -{ - smpi_process_init(&argc, &argv); - user_main_(); - return 0; -} - -inline static int smpi_main_wrapper(int argc, char **argv){ - int ret = smpi_simulated_main_(argc,argv); - if(ret !=0){ - XBT_WARN("SMPI process did not return 0. Return value : %d", ret); - smpi_process_data()->return_value=ret; - } - return 0; -} - -int __attribute__ ((weak)) main(int argc, char **argv) -{ - return smpi_main(smpi_main_wrapper, argc, argv); -} - -#endif - extern "C" { + static void smpi_init_logs(){ /* Connect log categories. See xbt/log.c */ @@ -703,78 +380,52 @@ static void smpi_init_logs(){ XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this function: xbt_log_appender_file.c depends on it DO NOT connect this in XBT or so, or it will be useless to xbt_log_appender_file.c */ XBT_LOG_CONNECT(instr_smpi); - XBT_LOG_CONNECT(smpi_base); XBT_LOG_CONNECT(smpi_bench); XBT_LOG_CONNECT(smpi_coll); XBT_LOG_CONNECT(smpi_colls); XBT_LOG_CONNECT(smpi_comm); + XBT_LOG_CONNECT(smpi_datatype); XBT_LOG_CONNECT(smpi_dvfs); XBT_LOG_CONNECT(smpi_group); XBT_LOG_CONNECT(smpi_kernel); XBT_LOG_CONNECT(smpi_mpi); - XBT_LOG_CONNECT(smpi_mpi_dt); + XBT_LOG_CONNECT(smpi_memory); + XBT_LOG_CONNECT(smpi_op); XBT_LOG_CONNECT(smpi_pmpi); + XBT_LOG_CONNECT(smpi_request); XBT_LOG_CONNECT(smpi_replay); XBT_LOG_CONNECT(smpi_rma); + XBT_LOG_CONNECT(smpi_shared); + XBT_LOG_CONNECT(smpi_utils); } } static void smpi_init_options(){ - int gather_id = find_coll_description(mpi_coll_gather_description, xbt_cfg_get_string("smpi/gather"),"gather"); - mpi_coll_gather_fun = reinterpret_cast - (mpi_coll_gather_description[gather_id].coll); - - int allgather_id = find_coll_description(mpi_coll_allgather_description, - xbt_cfg_get_string("smpi/allgather"),"allgather"); - mpi_coll_allgather_fun = reinterpret_cast - (mpi_coll_allgather_description[allgather_id].coll); - - int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description, - xbt_cfg_get_string("smpi/allgatherv"),"allgatherv"); - mpi_coll_allgatherv_fun = reinterpret_cast - (mpi_coll_allgatherv_description[allgatherv_id].coll); - - int allreduce_id = find_coll_description(mpi_coll_allreduce_description, - xbt_cfg_get_string("smpi/allreduce"),"allreduce"); - mpi_coll_allreduce_fun = reinterpret_cast - (mpi_coll_allreduce_description[allreduce_id].coll); - - int alltoall_id = find_coll_description(mpi_coll_alltoall_description, - xbt_cfg_get_string("smpi/alltoall"),"alltoall"); - mpi_coll_alltoall_fun = reinterpret_cast - (mpi_coll_alltoall_description[alltoall_id].coll); - - int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description, - xbt_cfg_get_string("smpi/alltoallv"),"alltoallv"); - mpi_coll_alltoallv_fun = reinterpret_cast - (mpi_coll_alltoallv_description[alltoallv_id].coll); - - int bcast_id = find_coll_description(mpi_coll_bcast_description, xbt_cfg_get_string("smpi/bcast"),"bcast"); - mpi_coll_bcast_fun = reinterpret_cast - (mpi_coll_bcast_description[bcast_id].coll); - - int reduce_id = find_coll_description(mpi_coll_reduce_description, xbt_cfg_get_string("smpi/reduce"),"reduce"); - mpi_coll_reduce_fun = reinterpret_cast - (mpi_coll_reduce_description[reduce_id].coll); - - int reduce_scatter_id = - find_coll_description(mpi_coll_reduce_scatter_description, - xbt_cfg_get_string("smpi/reduce-scatter"),"reduce_scatter"); - mpi_coll_reduce_scatter_fun = reinterpret_cast - (mpi_coll_reduce_scatter_description[reduce_scatter_id].coll); - - int scatter_id = find_coll_description(mpi_coll_scatter_description, xbt_cfg_get_string("smpi/scatter"),"scatter"); - mpi_coll_scatter_fun = reinterpret_cast - (mpi_coll_scatter_description[scatter_id].coll); - - int barrier_id = find_coll_description(mpi_coll_barrier_description, xbt_cfg_get_string("smpi/barrier"),"barrier"); - mpi_coll_barrier_fun = reinterpret_cast - (mpi_coll_barrier_description[barrier_id].coll); - - smpi_coll_cleanup_callback=nullptr; + + 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"); - smpi_privatize_global_variables = xbt_cfg_get_boolean("smpi/privatize-global-variables"); + const char* smpi_privatize_option = xbt_cfg_get_string("smpi/privatize-global-variables"); + 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/privatize-global-variables: %s", + smpi_privatize_option); + if (smpi_cpu_threshold < 0) smpi_cpu_threshold = DBL_MAX; @@ -791,7 +442,61 @@ static void smpi_init_options(){ } } -int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[]) +static int execute_command(const char * const argv[]) +{ + pid_t pid; + int status; + if (posix_spawnp(&pid, argv[0], nullptr, nullptr, (char* const*) argv, environ) != 0) + return 127; + if (waitpid(pid, &status, 0) != pid) + return 127; + return status; +} + +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)(void); + +static int smpi_run_entry_point(smpi_entry_point_type entry_point, std::vector args) +{ + const int argc = args.size(); + std::unique_ptr argv(new char*[argc + 1]); + for (int i = 0; i != argc; ++i) + argv[i] = args[i].empty() ? const_cast(""): &args[i].front(); + argv[argc] = nullptr; + + int res = entry_point(argc, argv.get()); + if (res != 0){ + XBT_WARN("SMPI process did not return 0. Return value : %d", res); + smpi_process()->set_return_value(res); + } + return 0; +} + +// TODO, remove the number of functions involved here +static smpi_entry_point_type smpi_resolve_function(void* handle) +{ + smpi_fortran_entry_point_type entry_point2 = + (smpi_fortran_entry_point_type) dlsym(handle, "user_main_"); + if (entry_point2 != nullptr) { + // fprintf(stderr, "EP user_main_=%p\n", entry_point2); + return [entry_point2](int argc, char** argv) { + smpi_process_init(&argc, &argv); + entry_point2(); + return 0; + }; + } + + smpi_c_entry_point_type entry_point = (smpi_c_entry_point_type) dlsym(handle, "main"); + if (entry_point != nullptr) { + // fprintf(stderr, "EP main=%p\n", entry_point); + return entry_point; + } + + return smpi_entry_point_type(); +} + +int smpi_main(const char* executable, int argc, char *argv[]) { srand(SMPI_RAND_SEED); @@ -800,31 +505,108 @@ int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[]) * configuration tools */ return 0; } - smpi_init_logs(); TRACE_global_init(&argc, argv); - TRACE_add_start_function(TRACE_smpi_alloc); - TRACE_add_end_function(TRACE_smpi_release); SIMIX_global_init(&argc, argv); MSG_init(&argc,argv); SMPI_switch_data_segment = &smpi_switch_data_segment; - smpi_init_options(); + simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) { + host.extension_set(new simgrid::smpi::SmpiHost(&host)); + }); // parse the platform file: get the host list SIMIX_create_environment(argv[1]); - SIMIX_comm_set_copy_data_callback(smpi_comm_copy_data_callback); - SIMIX_function_register_default(realmain); - SIMIX_launch_application(argv[2]); + SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback); - smpi_global_init(); + static std::size_t rank = 0; - smpi_check_options(); + if (smpi_privatize_global_variables == SMPI_PRIVATIZE_DLOPEN) { - if(smpi_privatize_global_variables) - smpi_initialize_global_memory_segments(); + std::string executable_copy = executable; + + // Prepare the copy of the binary (open the file and get its size) + // (fdin will remain open for the whole process execution. That's a sort of leak but we can live with it) + int fdin = open(executable_copy.c_str(), O_RDONLY); + xbt_assert(fdin >= 0, "Cannot read from %s", executable_copy.c_str()); + struct stat fdin_stat; + fstat(fdin, &fdin_stat); + off_t fdin_size = fdin_stat.st_size; + + simix_global->default_function = [executable_copy, fdin, fdin_size](std::vector args) { + return std::function([executable_copy, fdin, fdin_size, args] { + + // Copy the dynamic library: + std::string target_executable = executable_copy + + "_" + std::to_string(getpid()) + + "_" + std::to_string(rank++) + ".so"; + + int fdout = open(target_executable.c_str(), O_WRONLY); + xbt_assert(fdout >= 0, "Cannot write into %s", target_executable.c_str()); + +#if HAVE_SENDFILE + sendfile(fdout, fdin, NULL, fdin_size); +#else + const int bufsize = 1024 * 1024 * 4; + char buf[bufsize]; + while (int got = read(fdin, buf, bufsize)) { + if (got == -1) { + xbt_assert(errno == EINTR, "Cannot read from %s", executable_copy.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_executable.c_str()); + } else { + p += done; + todo -= done; + } + } + } + } +#endif + close(fdout); + + // Load the copy and resolve the entry point: + void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND); + unlink(target_executable.c_str()); + if (handle == nullptr) + xbt_die("dlopen failed"); + smpi_entry_point_type entry_point = smpi_resolve_function(handle); + if (!entry_point) + xbt_die("Could not resolve entry point"); + + smpi_run_entry_point(entry_point, args); + }); + }; + + } + else { + + // Load the dynamic library and resolve the entry point: + void* handle = dlopen(executable, RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND); + if (handle == nullptr) + xbt_die("dlopen failed for %s", executable); + smpi_entry_point_type entry_point = smpi_resolve_function(handle); + if (!entry_point) + xbt_die("main not found in %s", executable); + // TODO, register the executable for SMPI privatization + + // Execute the same entry point for each simulated process: + simix_global->default_function = [entry_point](std::vector args) { + return std::function([entry_point, args] { + smpi_run_entry_point(entry_point, args); + }); + }; + + } + + SIMIX_launch_application(argv[2]); + + SMPI_init(); /* Clean IO before the run */ fflush(stdout); @@ -852,8 +634,8 @@ int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[]) int count = smpi_process_count(); int i, ret=0; for (i = 0; i < count; i++) { - if(process_data[i]->return_value!=0){ - ret=process_data[i]->return_value;//return first non 0 value + if(process_data[i]->return_value()!=0){ + ret=process_data[i]->return_value();//return first non 0 value break; } } @@ -864,19 +646,44 @@ int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[]) return ret; } -// This function can be called from extern file, to initialize logs, options, and processes of smpi -// without the need of smpirun +// Called either directly from the user code, or from the code called by smpirun void SMPI_init(){ smpi_init_logs(); smpi_init_options(); smpi_global_init(); smpi_check_options(); - if (TRACE_is_enabled() && TRACE_is_configured()) - TRACE_smpi_alloc(); - if(smpi_privatize_global_variables) + TRACE_smpi_alloc(); + simgrid::surf::surfExitCallbacks.connect(TRACE_smpi_release); + if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) smpi_initialize_global_memory_segments(); } void SMPI_finalize(){ smpi_global_destroy(); } + +void smpi_mpi_init() { + if(smpi_init_sleep > 0) + simcall_process_sleep(smpi_init_sleep); +} + +double smpi_mpi_wtime(){ + double time; + if (smpi_process()->initialized() != 0 && smpi_process()->finalized() == 0 && smpi_process()->sampling() == 0) { + smpi_bench_end(); + time = SIMIX_get_clock(); + // to avoid deadlocks if used as a break condition, such as + // while (MPI_Wtime(...) < time_limit) { + // .... + // } + // because the time will not normally advance when only calls to MPI_Wtime + // are made -> deadlock (MPI_Wtime never reaches the time limit) + if(smpi_wtime_sleep > 0) + simcall_process_sleep(smpi_wtime_sleep); + smpi_bench_begin(); + } else { + time = SIMIX_get_clock(); + } + return time; +} +