From: Christian Heinrich Date: Fri, 3 Nov 2017 15:56:17 +0000 (+0100) Subject: [SMPI] Use SMPI-Instances also when smpirun is used. X-Git-Tag: v3.18~305^2~2 X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/commitdiff_plain/1cf17e3db34c0d42d051b7deb70c3aae28417d4f [SMPI] Use SMPI-Instances also when smpirun is used. This commit ensures that smpirun is not really a special case any more; freeing memory etc. is now all done by the SMPI instance. This also removes any assumption on where the finalization barrier for the process is stored; it's now always within the instance. --- diff --git a/src/smpi/internals/smpi_deployment.cpp b/src/smpi/internals/smpi_deployment.cpp index 407fe67112..d5ef6a582b 100644 --- a/src/smpi/internals/smpi_deployment.cpp +++ b/src/smpi/internals/smpi_deployment.cpp @@ -56,7 +56,9 @@ extern int* index_to_process_data; */ void SMPI_app_instance_register(const char *name, xbt_main_func_t code, int num_processes) { - SIMIX_function_register(name, code); + if (code != nullptr) { // When started with smpirun, we will not execute a function + SIMIX_function_register(name, code); + } static int already_called = 0; if (not already_called) { @@ -68,6 +70,9 @@ void SMPI_app_instance_register(const char *name, xbt_main_func_t code, int num_ } Instance instance(name, num_processes, process_count, MPI_COMM_NULL, MSG_barrier_init(num_processes)); + MPI_Group group = new simgrid::smpi::Group(instance.size); + instance.comm_world = new simgrid::smpi::Comm(group, nullptr); + MPI_Attr_put(instance.comm_world, MPI_UNIVERSE_SIZE, reinterpret_cast(instance.size)); process_count+=num_processes; @@ -84,10 +89,6 @@ void smpi_deployment_register_process(const char* instance_id, int rank, int ind Instance& instance = smpi_instances.at(instance_id); - if (instance.comm_world == MPI_COMM_NULL) { - MPI_Group group = new simgrid::smpi::Group(instance.size); - instance.comm_world = new simgrid::smpi::Comm(group, nullptr); - } instance.present_processes++; index_to_process_data[index] = instance.index + rank; instance.comm_world->group()->set_mapping(index, rank); @@ -115,9 +116,8 @@ msg_bar_t smpi_deployment_finalization_barrier(const char* instance_id) void smpi_deployment_cleanup_instances(){ for (auto const& item : smpi_instances) { Instance instance = item.second; - if (instance.comm_world != MPI_COMM_NULL) - delete instance.comm_world->group(); - delete instance.comm_world; MSG_barrier_destroy(instance.finalization_barrier); + simgrid::smpi::Comm::destroy(instance.comm_world); } + smpi_instances.clear(); } diff --git a/src/smpi/internals/smpi_global.cpp b/src/smpi/internals/smpi_global.cpp index 6d71a0c26b..351a95edee 100644 --- a/src/smpi/internals/smpi_global.cpp +++ b/src/smpi/internals/smpi_global.cpp @@ -58,10 +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( @@ -327,28 +340,29 @@ 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 - MPI_Group group; - 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); + } } } @@ -356,12 +370,7 @@ void smpi_global_destroy() { 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(); - } + 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()); @@ -376,13 +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;