X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/2b2677ab2f08298e2517f5ee3f56dfb9916234b6..9755bcbdab9bd6813d914915ef3887b78166f845:/src/smpi/mpi/smpi_comm.cpp diff --git a/src/smpi/mpi/smpi_comm.cpp b/src/smpi/mpi/smpi_comm.cpp index f3cec23338..e4e9929a0b 100644 --- a/src/smpi/mpi/smpi_comm.cpp +++ b/src/smpi/mpi/smpi_comm.cpp @@ -28,7 +28,7 @@ namespace smpi{ std::unordered_map Comm::keyvals_; int Comm::keyval_id_=0; -Comm::Comm(MPI_Group group, MPI_Topology topo, int smp) : group_(group), topo_(topo),is_smp_comm_(smp) +Comm::Comm(MPI_Group group, MPI_Topology topo, int smp, int in_id) : group_(group), topo_(topo),is_smp_comm_(smp), id_(in_id) { refcount_ = 1; topoType_ = MPI_INVALID_TOPO; @@ -38,6 +38,22 @@ Comm::Comm(MPI_Group group, MPI_Topology topo, int smp) : group_(group), topo_(t non_uniform_map_ = nullptr; leaders_map_ = nullptr; is_blocked_ = 0; + info_ = MPI_INFO_NULL; + errhandler_ = MPI_ERRORS_ARE_FATAL; + errhandler_->ref(); + //First creation of comm is done before SIMIX_run, so only do comms for others + if(in_id==MPI_UNDEFINED && smp==0 && this->rank()!=MPI_UNDEFINED ){ + int id; + if(this->rank()==0){ + static int global_id_ = 0; + id=global_id_; + global_id_++; + } + colls::bcast(&id, 1, MPI_INT, 0, this); + XBT_DEBUG("Communicator %p has id %d", this, id); + id_=id;//only set here, as we don't want to change it in the middle of the bcast + colls::barrier(this); + } } void Comm::destroy(Comm* comm) @@ -46,12 +62,11 @@ void Comm::destroy(Comm* comm) Comm::destroy(smpi_process()->comm_world()); return; } - delete comm->topo_; // there's no use count on topos Comm::unref(comm); } int Comm::dup(MPI_Comm* newcomm){ - if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) { + if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) { // we need to switch as the called function may silently touch global variables smpi_switch_data_segment(s4u::Actor::self()); } @@ -89,6 +104,26 @@ int Comm::dup(MPI_Comm* newcomm){ } } } + //duplicate info if present + if(info_!=MPI_INFO_NULL) + (*newcomm)->info_ = new simgrid::smpi::Info(info_); + //duplicate errhandler + (*newcomm)->set_errhandler(errhandler_); + return ret; +} + +int Comm::dup_with_info(MPI_Info info, MPI_Comm* newcomm){ + int ret = dup(newcomm); + if(ret != MPI_SUCCESS) + return ret; + if((*newcomm)->info_!=MPI_INFO_NULL){ + simgrid::smpi::Info::unref((*newcomm)->info_); + (*newcomm)->info_=MPI_INFO_NULL; + } + if(info != MPI_INFO_NULL){ + info->ref(); + (*newcomm)->info_=info; + } return ret; } @@ -113,6 +148,11 @@ int Comm::rank() return group_->rank(s4u::Actor::self()); } +int Comm::id() +{ + return id_; +} + void Comm::get_name (char* name, int* len) { if (this == MPI_COMM_UNINITIALIZED){ @@ -130,7 +170,7 @@ void Comm::get_name (char* name, int* len) } } -void Comm::set_name (char* name) +void Comm::set_name (const char* name) { if (this == MPI_COMM_UNINITIALIZED){ smpi_process()->comm_world()->set_name(name); @@ -194,7 +234,7 @@ MPI_Comm Comm::split(int color, int key) { if (this == MPI_COMM_UNINITIALIZED) return smpi_process()->comm_world()->split(color, key); - int system_tag = 123; + int system_tag = -123; int* recvbuf; MPI_Group group_root = nullptr; @@ -211,7 +251,7 @@ MPI_Comm Comm::split(int color, int key) } else { recvbuf = nullptr; } - Coll_gather_default::gather(sendbuf, 2, MPI_INT, recvbuf, 2, MPI_INT, 0, this); + gather__default(sendbuf, 2, MPI_INT, recvbuf, 2, MPI_INT, 0, this); xbt_free(sendbuf); /* Do the actual job */ if (myrank == 0) { @@ -236,7 +276,7 @@ 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::ActorPtr actor = group->actor(rankmap[j].second); + s4u::Actor* actor = group->actor(rankmap[j].second); group_out->set_mapping(actor, j); } MPI_Request* requests = xbt_new(MPI_Request, rankmap.size()); @@ -295,13 +335,44 @@ void Comm::unref(Comm* comm){ if(comm->refcount_==0){ comm->cleanup_smp(); comm->cleanup_attr(); + if (comm->info_ != MPI_INFO_NULL) + simgrid::smpi::Info::unref(comm->info_); + if (comm->errhandler_ != MPI_ERRHANDLER_NULL) + simgrid::smpi::Errhandler::unref(comm->errhandler_); + delete comm->topo_; // there's no use count on topos 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.ciface()) != 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); + MPI_Group group_intra = new Group(intra_comm_size); + int i = 0; + for (auto& actor : actor_list) { + if (this->group()->rank(actor.ciface()) != MPI_UNDEFINED) { + group_intra->set_mapping(actor.ciface(), i); + i++; + } + } + *leader=min_index; + return new Comm(group_intra, nullptr, 1); +} + void Comm::init_smp(){ int leader = -1; - + int i = 0; if (this == MPI_COMM_UNINITIALIZED) smpi_process()->comm_world()->init_smp(); @@ -315,44 +386,22 @@ void Comm::init_smp(){ smpi_process()->set_replaying(false); } - if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) { + if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) { // we need to switch as the called function may silently touch global variables smpi_switch_data_segment(s4u::Actor::self()); } - //identify neighbours in comm - //get the indices of all processes sharing the same simix host - auto& process_list = sg_host_self()->pimpl_->process_list_; - int intra_comm_size = 0; - int min_index = INT_MAX; // the minimum index will be the leader - for (auto& actor : process_list) { - int index = actor.get_pid(); - if (this->group()->rank(actor.iface()) != 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); - MPI_Group group_intra = new Group(intra_comm_size); - int i = 0; - for (auto& actor : process_list) { - if (this->group()->rank(actor.iface()) != MPI_UNDEFINED) { - group_intra->set_mapping(actor.iface(), i); - i++; - } - } + // identify neighbors in comm + MPI_Comm comm_intra = find_intra_comm(&leader); - MPI_Comm comm_intra = new Comm(group_intra, nullptr, 1); - leader=min_index; int* leaders_map = new int[comm_size]; int* leader_list = new int[comm_size]; std::fill_n(leaders_map, comm_size, 0); std::fill_n(leader_list, comm_size, -1); - Coll_allgather_ring::allgather(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, this); + allgather__ring(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, this); - if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) { + if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) { // we need to switch as the called function may silently touch global variables smpi_switch_data_segment(s4u::Actor::self()); } @@ -375,6 +424,7 @@ void Comm::init_smp(){ leader_group_size++; } } + xbt_assert(leader_group_size > 0); std::sort(leader_list, leader_list + leader_group_size); MPI_Group leaders_group = new Group(leader_group_size); @@ -383,7 +433,7 @@ 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]), i); + leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]).get(), i); leader_comm = new Comm(leaders_group, nullptr,1); this->set_leaders_comm(leader_comm); this->set_intra_comm(comm_intra); @@ -391,7 +441,7 @@ 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]), i); + leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]).get(), i); if(this->get_leaders_comm()==MPI_COMM_NULL){ leader_comm = new Comm(leaders_group, nullptr,1); @@ -408,7 +458,7 @@ void Comm::init_smp(){ if(comm_intra->rank()==0) { int is_uniform = 1; int* non_uniform_map = xbt_new0(int,leader_group_size); - Coll_allgather_ring::allgather(&my_local_size, 1, MPI_INT, + allgather__ring(&my_local_size, 1, MPI_INT, non_uniform_map, 1, MPI_INT, leader_comm); for(i=0; i < leader_group_size; i++) { if(non_uniform_map[0] != non_uniform_map[i]) { @@ -423,9 +473,9 @@ void Comm::init_smp(){ } is_uniform_=is_uniform; } - Coll_bcast_scatter_LR_allgather::bcast(&(is_uniform_),1, MPI_INT, 0, comm_intra ); + bcast__scatter_LR_allgather(&(is_uniform_),1, MPI_INT, 0, comm_intra ); - if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) { + if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) { // we need to switch as the called function may silently touch global variables smpi_switch_data_segment(s4u::Actor::self()); } @@ -442,7 +492,7 @@ void Comm::init_smp(){ } int global_blocked; - Coll_allreduce_default::allreduce(&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){ @@ -465,7 +515,7 @@ MPI_Comm Comm::f2c(int id) { } else if(F2C::f2c_lookup() != nullptr && id >= 0) { char key[KEY_SIZE]; const auto& lookup = F2C::f2c_lookup(); - auto comm = lookup->find(get_key_id(key, id)); + auto comm = lookup->find(get_key(key, id)); return comm == lookup->end() ? MPI_COMM_NULL : static_cast(comm->second); } else { return MPI_COMM_NULL; @@ -474,17 +524,7 @@ MPI_Comm Comm::f2c(int id) { void Comm::free_f(int id) { char key[KEY_SIZE]; - F2C::f2c_lookup()->erase(id == 0 ? get_key(key, id) : get_key_id(key, id)); -} - -int Comm::add_f() { - if(F2C::f2c_lookup()==nullptr){ - F2C::set_f2c_lookup(new std::unordered_map); - } - char key[KEY_SIZE]; - (*(F2C::f2c_lookup()))[this == MPI_COMM_WORLD ? get_key(key, F2C::f2c_id()) : get_key_id(key, F2C::f2c_id())] = this; - f2c_id_increment(); - return F2C::f2c_id()-1; + F2C::f2c_lookup()->erase(get_key(key, id)); } void Comm::add_rma_win(MPI_Win win){ @@ -504,28 +544,54 @@ void Comm::finish_rma_calls(){ } } -MPI_Info Comm::info(){ - if(info_== MPI_INFO_NULL) +MPI_Info Comm::info() +{ + if (info_ == MPI_INFO_NULL) info_ = new Info(); info_->ref(); return info_; } -void Comm::set_info(MPI_Info info){ - if(info_!= MPI_INFO_NULL) +void Comm::set_info(MPI_Info info) +{ + if (info_ != MPI_INFO_NULL) + simgrid::smpi::Info::unref(info); + info_ = info; + if (info_ != MPI_INFO_NULL) info->ref(); - info_=info; +} + +MPI_Errhandler Comm::errhandler() +{ + if (errhandler_ != MPI_ERRHANDLER_NULL) + errhandler_->ref(); + return errhandler_; +} + +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(); } MPI_Comm Comm::split_type(int type, int /*key*/, MPI_Info) { - if(type != MPI_COMM_TYPE_SHARED){ + //MPI_UNDEFINED can be given to some nodes... but we need them to still perform the smp part which is collective + if(type != MPI_COMM_TYPE_SHARED && type != MPI_UNDEFINED){ + return MPI_COMM_NULL; + } + int leader=0; + MPI_Comm res= this->find_intra_comm(&leader); + if(type != MPI_UNDEFINED) + return res; + else{ + xbt_assert(res->refcount_ == 1); // ensure the next call to Comm::destroy really frees the comm + Comm::destroy(res); return MPI_COMM_NULL; } - this->init_smp(); - this->ref(); - this->get_intra_comm()->ref(); - return this->get_intra_comm(); } } // namespace smpi