X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/67fe7b9d6c00b390a8598bc1e72d42d8343cb218..15f804cfddebc4fe4b305acd1c66487f2405420d:/src/smpi/mpi/smpi_comm.cpp diff --git a/src/smpi/mpi/smpi_comm.cpp b/src/smpi/mpi/smpi_comm.cpp index d4ef0530d0..d8ca45b523 100644 --- a/src/smpi/mpi/smpi_comm.cpp +++ b/src/smpi/mpi/smpi_comm.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2010-2018. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2010-2019. 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. */ @@ -8,6 +8,7 @@ #include "smpi_datatype.hpp" #include "smpi_request.hpp" #include "smpi_win.hpp" +#include "smpi_info.hpp" #include "src/smpi/include/smpi_actor.hpp" #include "src/surf/HostImpl.hpp" @@ -18,8 +19,6 @@ 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; -using simgrid::s4u::ActorPtr; - /* 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. */ @@ -29,7 +28,7 @@ namespace smpi{ std::unordered_map Comm::keyvals_; int Comm::keyval_id_=0; -Comm::Comm(MPI_Group group, MPI_Topology topo) : group_(group), topo_(topo) +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; @@ -39,6 +38,21 @@ Comm::Comm(MPI_Group group, MPI_Topology topo) : group_(group), topo_(topo) non_uniform_map_ = nullptr; leaders_map_ = nullptr; is_blocked_ = 0; + info_ = MPI_INFO_NULL; + errhandler_ = MPI_ERRORS_ARE_FATAL; + //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) @@ -47,38 +61,67 @@ 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(simgrid::s4u::Actor::self()); + smpi_switch_data_segment(s4u::Actor::self()); } MPI_Group cp = new Group(this->group()); (*newcomm) = new Comm(cp, this->topo()); int ret = MPI_SUCCESS; if (not attributes()->empty()) { - int flag; - void* value_out; + int flag=0; + void* value_out=nullptr; for (auto const& it : *attributes()) { smpi_key_elem elem = keyvals_.at(it.first); - if (elem != nullptr && elem->copy_fn.comm_copy_fn != MPI_NULL_COPY_FN) { - ret = elem->copy_fn.comm_copy_fn(this, it.first, nullptr, it.second, &value_out, &flag); + 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 (flag){ + 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}); } } } } + //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){ + (*newcomm)->info_=info; + } return ret; } @@ -89,10 +132,6 @@ MPI_Group Comm::group() return group_; } -MPI_Topology Comm::topo() { - return topo_; -} - int Comm::size() { if (this == MPI_COMM_UNINITIALIZED) @@ -104,7 +143,12 @@ int Comm::rank() { if (this == MPI_COMM_UNINITIALIZED) return smpi_process()->comm_world()->rank(); - return group_->rank(simgrid::s4u::Actor::self()); + return group_->rank(s4u::Actor::self()); +} + +int Comm::id() +{ + return id_; } void Comm::get_name (char* name, int* len) @@ -113,14 +157,27 @@ void Comm::get_name (char* name, int* len) smpi_process()->comm_world()->get_name(name, len); return; } - if(this == MPI_COMM_WORLD) { - strncpy(name, "WORLD", 6); - *len = 5; + 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, "%p", this); + *len = snprintf(name, MPI_MAX_NAME_STRING+1, "%s", name_.c_str()); + } +} + +void Comm::set_name (const char* name) +{ + if (this == MPI_COMM_UNINITIALIZED){ + smpi_process()->comm_world()->set_name(name); + return; } + name_.replace (0, MPI_MAX_NAME_STRING+1, name); } + void Comm::set_leaders_comm(MPI_Comm leaders){ if (this == MPI_COMM_UNINITIALIZED){ smpi_process()->comm_world()->set_leaders_comm(leaders); @@ -129,10 +186,6 @@ void Comm::set_leaders_comm(MPI_Comm leaders){ leaders_comm_=leaders; } -void Comm::set_intra_comm(MPI_Comm leaders){ - intra_comm_=leaders; -} - int* Comm::get_non_uniform_map(){ if (this == MPI_COMM_UNINITIALIZED) return smpi_process()->comm_world()->get_non_uniform_map(); @@ -169,31 +222,37 @@ int Comm::is_blocked(){ return is_blocked_; } +int Comm::is_smp_comm(){ + if (this == MPI_COMM_UNINITIALIZED) + return smpi_process()->comm_world()->is_smp_comm(); + return is_smp_comm_; +} + 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; MPI_Group group_out = nullptr; MPI_Group group = this->group(); - int rank = this->rank(); + int myrank = this->rank(); int size = this->size(); /* Gather all colors and keys on rank 0 */ int* sendbuf = xbt_new(int, 2); sendbuf[0] = color; sendbuf[1] = key; - if(rank == 0) { + if (myrank == 0) { recvbuf = xbt_new(int, 2 * size); } 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(rank == 0) { + if (myrank == 0) { MPI_Group* group_snd = xbt_new(MPI_Group, size); std::vector> rankmap; rankmap.reserve(size); @@ -215,7 +274,7 @@ MPI_Comm Comm::split(int color, int key) group_root = group_out; /* Save root's group */ } for (unsigned j = 0; j < rankmap.size(); j++) { - 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()); @@ -274,13 +333,40 @@ void Comm::unref(Comm* comm){ if(comm->refcount_==0){ comm->cleanup_smp(); comm->cleanup_attr(); + 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(); @@ -294,46 +380,24 @@ 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(simgrid::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.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++; - } + smpi_switch_data_segment(s4u::Actor::self()); } + // identify neighbors in comm + MPI_Comm comm_intra = find_intra_comm(&leader); - MPI_Comm comm_intra = new Comm(group_intra, nullptr); - 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_mpich::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(simgrid::s4u::Actor::self()); + smpi_switch_data_segment(s4u::Actor::self()); } if(leaders_map_==nullptr){ @@ -362,18 +426,18 @@ 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(simgrid::s4u::Actor::by_pid(leader_list[i]), i); - leader_comm = new Comm(leaders_group, nullptr); + 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); // create intracommunicator }else{ for (i=0; i< leader_group_size;i++) - leaders_group->set_mapping(simgrid::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); + leader_comm = new Comm(leaders_group, nullptr,1); this->set_leaders_comm(leader_comm); }else{ leader_comm=this->get_leaders_comm(); @@ -387,7 +451,7 @@ 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_mpich::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]) { @@ -402,11 +466,11 @@ void Comm::init_smp(){ } is_uniform_=is_uniform; } - Coll_bcast_mpich::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(simgrid::s4u::Actor::self()); + smpi_switch_data_segment(s4u::Actor::self()); } // Are the ranks blocked ? = allocated contiguously on the SMP nodes int is_blocked=1; @@ -421,7 +485,7 @@ 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){ @@ -444,7 +508,7 @@ 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; @@ -453,17 +517,7 @@ 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){ @@ -483,18 +537,45 @@ void Comm::finish_rma_calls(){ } } -MPI_Comm Comm::split_type(int type, int key, MPI_Info info) -{ - if(type != MPI_COMM_TYPE_SHARED){ - return MPI_COMM_NULL; - } - this->init_smp(); - this->ref(); - this->get_intra_comm()->ref(); - return this->get_intra_comm(); +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) + info->ref(); + info_=info; } + +MPI_Errhandler Comm::errhandler(){ + return errhandler_; +} + +void Comm::set_errhandler(MPI_Errhandler errhandler){ + errhandler_=errhandler; + if(errhandler_!= MPI_ERRHANDLER_NULL) + errhandler->ref(); } +MPI_Comm Comm::split_type(int type, int /*key*/, MPI_Info) +{ + //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; + } +} +} // namespace smpi +} // namespace simgrid