X-Git-Url: http://bilbo.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/331be63536e6ac6588c60e66e27dede80da9cd8d..7f671da878195261904d4bbc180a697902799dad:/src/smpi/mpi/smpi_comm.cpp diff --git a/src/smpi/mpi/smpi_comm.cpp b/src/smpi/mpi/smpi_comm.cpp index 6ed5725584..829cf51b1f 100644 --- a/src/smpi/mpi/smpi_comm.cpp +++ b/src/smpi/mpi/smpi_comm.cpp @@ -12,12 +12,22 @@ #include "src/smpi/include/smpi_actor.hpp" #include "src/surf/HostImpl.hpp" -#include +#include XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_comm, smpi, "Logging specific to SMPI (comm)"); -simgrid::smpi::Comm mpi_MPI_COMM_UNINITIALIZED; -MPI_Comm MPI_COMM_UNINITIALIZED=&mpi_MPI_COMM_UNINITIALIZED; +simgrid::smpi::Comm smpi_MPI_COMM_UNINITIALIZED; +MPI_Comm MPI_COMM_UNINITIALIZED=&smpi_MPI_COMM_UNINITIALIZED; +/** + * 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 (the one their SMPI instance uses). + * + */ +MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED; /* Support for cartesian topology was added, but there are 2 other types of topology, graph et dist graph. In order to * support them, we have to add a field SMPI_Topo_type, and replace the MPI_Topology field by an union. */ @@ -55,6 +65,8 @@ void Comm::destroy(Comm* comm) Comm::destroy(smpi_process()->comm_world()); return; } + if(comm != MPI_COMM_WORLD) + comm->mark_as_deleted(); Comm::unref(comm); } @@ -67,33 +79,29 @@ int Comm::dup(MPI_Comm* newcomm){ (*newcomm) = new Comm(cp, this->topo()); int ret = MPI_SUCCESS; - if (not attributes()->empty()) { - int flag=0; - void* value_out=nullptr; - for (auto const& it : *attributes()) { - smpi_key_elem elem = keyvals_.at(it.first); - if (elem != nullptr){ - if( elem->copy_fn.comm_copy_fn != MPI_NULL_COPY_FN && - elem->copy_fn.comm_copy_fn != MPI_COMM_DUP_FN) - ret = elem->copy_fn.comm_copy_fn(this, it.first, elem->extra_state, it.second, &value_out, &flag); - else if ( elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN && - *(int*)*elem->copy_fn.comm_copy_fn_fort != 1){ - value_out=(int*)xbt_malloc(sizeof(int)); - elem->copy_fn.comm_copy_fn_fort(this, it.first, elem->extra_state, it.second, value_out, &flag,&ret); - } - if (ret != MPI_SUCCESS) { - Comm::destroy(*newcomm); - *newcomm = MPI_COMM_NULL; - return ret; - } - if (elem->copy_fn.comm_copy_fn == MPI_COMM_DUP_FN || - ((elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN) && *(int*)*elem->copy_fn.comm_copy_fn_fort == 1)){ - elem->refcount++; - (*newcomm)->attributes()->insert({it.first, it.second}); - }else if (flag){ - elem->refcount++; - (*newcomm)->attributes()->insert({it.first, value_out}); - } + for (auto const& it : attributes()) { + smpi_key_elem elem = keyvals_.at(it.first); + if (elem != nullptr) { + int flag = 0; + void* value_out = nullptr; + if (elem->copy_fn.comm_copy_fn != MPI_NULL_COPY_FN && elem->copy_fn.comm_copy_fn != MPI_COMM_DUP_FN) + ret = elem->copy_fn.comm_copy_fn(this, it.first, elem->extra_state, it.second, &value_out, &flag); + else if (elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN && *(int*)*elem->copy_fn.comm_copy_fn_fort != 1) { + value_out = (int*)xbt_malloc(sizeof(int)); + elem->copy_fn.comm_copy_fn_fort(this, it.first, elem->extra_state, it.second, value_out, &flag, &ret); + } + if (ret != MPI_SUCCESS) { + Comm::destroy(*newcomm); + *newcomm = MPI_COMM_NULL; + return ret; + } + if (elem->copy_fn.comm_copy_fn == MPI_COMM_DUP_FN || + ((elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN) && *(int*)*elem->copy_fn.comm_copy_fn_fort == 1)) { + elem->refcount++; + (*newcomm)->attributes().emplace(it.first, it.second); + } else if (flag) { + elem->refcount++; + (*newcomm)->attributes().emplace(it.first, value_out); } } } @@ -101,7 +109,10 @@ int Comm::dup(MPI_Comm* newcomm){ if(info_!=MPI_INFO_NULL) (*newcomm)->info_ = new simgrid::smpi::Info(info_); //duplicate errhandler - (*newcomm)->set_errhandler(errhandler_); + if (errhandlers_ != nullptr)//MPI_COMM_WORLD, only grab our own + (*newcomm)->set_errhandler(errhandlers_[this->rank()]); + else + (*newcomm)->set_errhandler(errhandler_); return ret; } @@ -138,7 +149,7 @@ int Comm::rank() const { if (this == MPI_COMM_UNINITIALIZED) return smpi_process()->comm_world()->rank(); - return group_->rank(s4u::Actor::self()); + return group_->rank(s4u::this_actor::get_pid()); } int Comm::id() const @@ -155,14 +166,23 @@ void Comm::get_name(char* name, int* len) const if(this == MPI_COMM_WORLD && name_.empty()) { strncpy(name, "MPI_COMM_WORLD", 15); *len = 14; - } else if(this == MPI_COMM_SELF && name_.empty()) { - strncpy(name, "MPI_COMM_SELF", 14); - *len = 13; } else { *len = snprintf(name, MPI_MAX_NAME_STRING+1, "%s", name_.c_str()); } } +std::string Comm::name() const +{ + int size; + char name[MPI_MAX_NAME_STRING+1]; + this->get_name(name, &size); + if (name[0]=='\0') + return std::string("MPI_Comm"); + else + return std::string(name); +} + + void Comm::set_name (const char* name) { if (this == MPI_COMM_UNINITIALIZED){ @@ -238,7 +258,7 @@ MPI_Comm Comm::split(int color, int key) MPI_Group group_root = nullptr; MPI_Group group_out = nullptr; - MPI_Group group = this->group(); + const Group* group = this->group(); int myrank = this->rank(); int size = this->size(); /* Gather all colors and keys on rank 0 */ @@ -270,7 +290,7 @@ MPI_Comm Comm::split(int color, int key) group_root = group_out; /* Save root's group */ } for (unsigned j = 0; j < rankmap.size(); j++) { - s4u::Actor* actor = group->actor(rankmap[j].second); + aid_t actor = group->actor(rankmap[j].second); group_out->set_mapping(actor, j); } std::vector requests(rankmap.size()); @@ -321,43 +341,48 @@ void Comm::unref(Comm* comm){ return; } comm->refcount_--; - Group::unref(comm->group_); if(comm->refcount_==0){ if(simgrid::smpi::F2C::lookup() != nullptr) - F2C::free_f(comm->c2f()); + F2C::free_f(comm->f2c_id()); comm->cleanup_smp(); comm->cleanup_attr(); if (comm->info_ != MPI_INFO_NULL) simgrid::smpi::Info::unref(comm->info_); - if (comm->errhandler_ != MPI_ERRHANDLER_NULL) + if(comm->errhandlers_!=nullptr){ + for (int i=0; isize(); i++) + if (comm->errhandlers_[i]!=MPI_ERRHANDLER_NULL) + simgrid::smpi::Errhandler::unref(comm->errhandlers_[i]); + delete[] comm->errhandlers_; + } else if (comm->errhandler_ != MPI_ERRHANDLER_NULL) simgrid::smpi::Errhandler::unref(comm->errhandler_); - delete comm; } + Group::unref(comm->group_); + if(comm->refcount_==0) + delete comm; } MPI_Comm Comm::find_intra_comm(int * leader){ //get the indices of all processes sharing the same simix host - auto& actor_list = sg_host_self()->pimpl_->actor_list_; int intra_comm_size = 0; - int min_index = INT_MAX; // the minimum index will be the leader - for (auto& actor : actor_list) { - int index = actor.get_pid(); - if (this->group()->rank(actor.get_ciface()) != MPI_UNDEFINED) { // Is this process in the current group? + aid_t min_index = std::numeric_limits::max(); // the minimum index will be the leader + sg_host_self()->get_impl()->foreach_actor([this, &intra_comm_size, &min_index](auto& actor) { + aid_t index = actor.get_pid(); + if (this->group()->rank(index) != MPI_UNDEFINED) { // Is this process in the current group? intra_comm_size++; if (index < min_index) min_index = index; } - } + }); XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size); auto* group_intra = new Group(intra_comm_size); int i = 0; - for (auto& actor : actor_list) { - if (this->group()->rank(actor.get_ciface()) != MPI_UNDEFINED) { - group_intra->set_mapping(actor.get_ciface(), i); + sg_host_self()->get_impl()->foreach_actor([this, group_intra, &i](auto& actor) { + if (this->group()->rank(actor.get_pid()) != MPI_UNDEFINED) { + group_intra->set_mapping(actor.get_pid(), i); i++; } - } + }); *leader=min_index; return new Comm(group_intra, nullptr, true); } @@ -424,7 +449,7 @@ void Comm::init_smp(){ if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && this!=MPI_COMM_WORLD){ //create leader_communicator for (i=0; i< leader_group_size;i++) - leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]).get(), i); + leaders_group->set_mapping(leader_list[i], i); leader_comm = new Comm(leaders_group, nullptr, true); this->set_leaders_comm(leader_comm); this->set_intra_comm(comm_intra); @@ -432,7 +457,7 @@ void Comm::init_smp(){ // create intracommunicator }else{ for (i=0; i< leader_group_size;i++) - leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]).get(), i); + leaders_group->set_mapping(leader_list[i], i); if(this->get_leaders_comm()==MPI_COMM_NULL){ leader_comm = new Comm(leaders_group, nullptr, true); @@ -464,7 +489,7 @@ void Comm::init_smp(){ } is_uniform_=is_uniform; } - bcast__scatter_LR_allgather(&(is_uniform_),1, MPI_INT, 0, comm_intra ); + bcast__scatter_LR_allgather(&is_uniform_, 1, MPI_INT, 0, comm_intra); if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) { // we need to switch as the called function may silently touch global variables @@ -472,7 +497,7 @@ void Comm::init_smp(){ } // Are the ranks blocked ? = allocated contiguously on the SMP nodes int is_blocked=1; - int prev=this->group()->rank(comm_intra->group()->actor(0)); + int prev = this->group()->rank(comm_intra->group()->actor(0)); for (i = 1; i < my_local_size; i++) { int that = this->group()->rank(comm_intra->group()->actor(i)); if (that != prev + 1) { @@ -483,7 +508,7 @@ void Comm::init_smp(){ } int global_blocked; - allreduce__default(&is_blocked, &(global_blocked), 1, MPI_INT, MPI_LAND, this); + allreduce__default(&is_blocked, &global_blocked, 1, MPI_INT, MPI_LAND, this); if(MPI_COMM_WORLD==MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD){ if(this->rank()==0){ @@ -553,18 +578,36 @@ void Comm::set_info(MPI_Info info) MPI_Errhandler Comm::errhandler() { - if (errhandler_ != MPI_ERRHANDLER_NULL) - errhandler_->ref(); - return errhandler_; + if (this != MPI_COMM_WORLD){ + if (errhandler_ != MPI_ERRHANDLER_NULL) + errhandler_->ref(); + return errhandler_; + } else { + if(errhandlers_==nullptr) + return MPI_ERRORS_ARE_FATAL; + else { + if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL) + errhandlers_[this->rank()]->ref(); + return errhandlers_[this->rank()]; + } + } } void Comm::set_errhandler(MPI_Errhandler errhandler) { - if (errhandler_ != MPI_ERRHANDLER_NULL) - simgrid::smpi::Errhandler::unref(errhandler_); - errhandler_ = errhandler; - if (errhandler_ != MPI_ERRHANDLER_NULL) - errhandler_->ref(); + if(this != MPI_COMM_WORLD){ + if (errhandler_ != MPI_ERRHANDLER_NULL) + simgrid::smpi::Errhandler::unref(errhandler_); + errhandler_ = errhandler; + }else{ + if(errhandlers_==nullptr) + errhandlers_= new MPI_Errhandler[this->size()]{MPI_ERRHANDLER_NULL}; + if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL) + simgrid::smpi::Errhandler::unref(errhandlers_[this->rank()]); + errhandlers_[this->rank()]=errhandler; + } + if (errhandler != MPI_ERRHANDLER_NULL) + errhandler->ref(); } MPI_Comm Comm::split_type(int type, int /*key*/, const Info*)