X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/dbcf720995035476f8fda570bd5ab0926cacc799..a7c708d3570410ce5a1316a54f9dda55d2bb46ca:/src/smpi/internals/smpi_global.cpp diff --git a/src/smpi/internals/smpi_global.cpp b/src/smpi/internals/smpi_global.cpp index 70dc1e3406..f75d237792 100644 --- a/src/smpi/internals/smpi_global.cpp +++ b/src/smpi/internals/smpi_global.cpp @@ -6,8 +6,10 @@ #include "SmpiHost.hpp" #include "mc/mc.h" #include "private.hpp" +#include "simgrid/s4u/Engine.hpp" #include "simgrid/s4u/Host.hpp" #include "simgrid/s4u/Mailbox.hpp" +#include "simgrid/s4u/forward.hpp" #include "smpi_coll.hpp" #include "smpi_comm.hpp" #include "smpi_group.hpp" @@ -54,8 +56,9 @@ using simgrid::s4u::Actor; using simgrid::s4u::ActorPtr; std::unordered_map location2speedup; -static std::map process_data; +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; @@ -83,12 +86,6 @@ static simgrid::config::Flag smpi_init_sleep( void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t) = &smpi_comm_copy_buffer_callback; -void smpi_add_process(ActorPtr actor) -{ - process_data.insert( - {actor->getPid()-1, new simgrid::smpi::Process(actor, nullptr)}); -} - int smpi_process_count() { return process_count; @@ -103,9 +100,9 @@ simgrid::smpi::Process* smpi_process() return static_cast(msgExt->data); } -simgrid::smpi::Process* smpi_process_remote(int index) +simgrid::smpi::Process* smpi_process_remote(ActorPtr actor) { - return process_data.at(index); + return process_data.at(actor); } MPI_Comm smpi_process_comm_self(){ @@ -117,7 +114,7 @@ void smpi_process_init(int *argc, char ***argv){ } int smpi_process_index(){ - return smpi_process()->index(); + return simgrid::s4u::Actor::self()->getPid(); } void * smpi_process_get_user_data(){ @@ -195,10 +192,7 @@ void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t b 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()); + smpi_switch_data_segment(comm->src_proc->iface()); tmpbuff = static_cast(xbt_malloc(buff_size)); memcpy_private(tmpbuff, buff, private_blocks); } @@ -206,9 +200,7 @@ void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t b 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()); + smpi_switch_data_segment(comm->dst_proc->iface()); } XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff); memcpy_private(comm->dst_buff, tmpbuff, private_blocks); @@ -353,18 +345,6 @@ void smpi_global_destroy() smpi_bench_destroy(); smpi_shared_destroy(); smpi_deployment_cleanup_instances(); - for (auto& pair : process_data) { - auto& process = pair.second; - if (process->comm_self() != MPI_COMM_NULL) { - simgrid::smpi::Comm::destroy(process->comm_self()); - } - if (process->comm_intra() != MPI_COMM_NULL) { - simgrid::smpi::Comm::destroy(process->comm_intra()); - } - xbt_os_timer_free(process->timer()); - xbt_mutex_destroy(process->mailboxes_mutex()); - } - process_data.clear(); if (simgrid::smpi::Colls::smpi_coll_cleanup_callback != nullptr) simgrid::smpi::Colls::smpi_coll_cleanup_callback(); @@ -402,7 +382,7 @@ static void smpi_init_options(){ #if defined(__FreeBSD__) if (smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) { - XBT_INFO("Mixing mmap privatization is broken on FreeBSD, switching to dlopen privatization instead."); + XBT_INFO("mmap privatization is broken on FreeBSD, switching to dlopen privatization instead."); smpi_privatize_global_variables = SMPI_PRIVATIZE_DLOPEN; } #endif @@ -439,7 +419,8 @@ static int smpi_run_entry_point(smpi_entry_point_type entry_point, std::vectorset_return_value(res); + if (smpi_exit_status == 0) + smpi_exit_status = res; } return 0; } @@ -481,6 +462,9 @@ int smpi_main(const char* executable, int argc, char *argv[]) SMPI_switch_data_segment = &smpi_switch_data_segment; + // TODO This will not be executed in the case where smpi_main is not called, + // e.g., not for smpi_msg_masterslave. This should be moved to another location + // that is always called -- maybe close to Actor::onCreation? simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::SmpiHost(&host)); }); @@ -490,7 +474,6 @@ int smpi_main(const char* executable, int argc, char *argv[]) SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback); smpi_init_options(); - if (smpi_privatize_global_variables == SMPI_PRIVATIZE_DLOPEN) { std::string executable_copy = executable; @@ -510,7 +493,8 @@ int smpi_main(const char* executable, int argc, char *argv[]) + "_" + std::to_string(rank++) + ".so"; int fdin = open(executable_copy.c_str(), O_RDONLY); - xbt_assert(fdin >= 0, "Cannot read from %s", executable_copy.c_str()); + xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", + executable_copy.c_str()); int fdout = open(target_executable.c_str(), O_CREAT | O_RDWR, S_IRWXU); xbt_assert(fdout >= 0, "Cannot write into %s", target_executable.c_str()); @@ -562,7 +546,7 @@ int smpi_main(const char* executable, int argc, char *argv[]) else { // Load the dynamic library and resolve the entry point: - void* handle = dlopen(executable, RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND); + void* handle = dlopen(executable, RTLD_LAZY | RTLD_LOCAL); if (handle == nullptr) xbt_die("dlopen failed for %s: %s (errno: %d -- %s)", executable, dlerror(), errno, strerror(errno)); smpi_entry_point_type entry_point = smpi_resolve_function(handle); @@ -582,7 +566,7 @@ int smpi_main(const char* executable, int argc, char *argv[]) SMPI_init(); SIMIX_launch_application(argv[2]); SMPI_app_instance_register(smpi_default_instance_name, nullptr, - SIMIX_process_count()); // This call has a side effect on process_count... + 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; @@ -610,30 +594,33 @@ int smpi_main(const char* executable, int argc, char *argv[]) "You may want to use sampling functions or trace replay to reduce this."); } } - int ret = 0; - for (int i = 0, count = smpi_process_count(); i < count; i++) { - if (process_data.at(i)->return_value() != 0) { - ret = process_data.at(i)->return_value(); // return first non 0 value - break; - } - } smpi_global_destroy(); TRACE_end(); - return ret; + return smpi_exit_status; } // Called either directly from the user code, or from the code called by smpirun void SMPI_init(){ simgrid::s4u::Actor::onCreation.connect([](simgrid::s4u::ActorPtr actor) { - smpi_add_process(actor); + if (not actor->isDaemon()) { + process_data.insert({actor, new simgrid::smpi::Process(actor, nullptr)}); + } }); + simgrid::s4u::Actor::onDestruction.connect([](simgrid::s4u::ActorPtr actor) { + auto it = process_data.find(actor); + if (it != process_data.end()) { + delete it->second; + process_data.erase(it); + } + }); + smpi_init_options(); smpi_global_init(); smpi_check_options(); TRACE_smpi_alloc(); - simgrid::surf::surfExitCallbacks.connect(TRACE_smpi_release); + simgrid::s4u::onSimulationEnd.connect(TRACE_smpi_release); if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) smpi_backup_global_memory_segment(); }