X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/ae1d123508cf7ef1d7932ce0cecbd8945c062cfd..cf507642628052d29599e6b7703e049ef28c0505:/src/smpi/mpi/smpi_comm.cpp diff --git a/src/smpi/mpi/smpi_comm.cpp b/src/smpi/mpi/smpi_comm.cpp index ed178f0802..7722473d21 100644 --- a/src/smpi/mpi/smpi_comm.cpp +++ b/src/smpi/mpi/smpi_comm.cpp @@ -1,66 +1,43 @@ -/* Copyright (c) 2010-2017. 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. */ #include "smpi_comm.hpp" -#include "private.h" -#include "private.hpp" -#include "simgrid/s4u/Host.hpp" #include "smpi_coll.hpp" #include "smpi_datatype.hpp" -#include "smpi_process.hpp" #include "smpi_request.hpp" -#include "smpi_status.hpp" #include "smpi_win.hpp" -#include "src/simix/smx_private.h" -#include +#include "smpi_info.hpp" +#include "src/smpi/include/smpi_actor.hpp" +#include "src/surf/HostImpl.hpp" + #include XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_comm, smpi, "Logging specific to SMPI (comm)"); - simgrid::smpi::Comm mpi_MPI_COMM_UNINITIALIZED; +simgrid::smpi::Comm mpi_MPI_COMM_UNINITIALIZED; MPI_Comm MPI_COMM_UNINITIALIZED=&mpi_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. */ -static int smpi_compare_rankmap(const void *a, const void *b) -{ - const int* x = static_cast(a); - const int* y = static_cast(b); - - if (x[1] < y[1]) { - return -1; - } - if (x[1] == y[1]) { - if (x[0] < y[0]) { - return -1; - } - if (x[0] == y[0]) { - return 0; - } - return 1; - } - return 1; -} - namespace simgrid{ 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) : group_(group), topo_(topo),is_smp_comm_(smp) { - refcount_=1; - topoType_ = MPI_INVALID_TOPO; - intra_comm_ = MPI_COMM_NULL; - leaders_comm_ = MPI_COMM_NULL; - is_uniform_=1; + refcount_ = 1; + topoType_ = MPI_INVALID_TOPO; + intra_comm_ = MPI_COMM_NULL; + leaders_comm_ = MPI_COMM_NULL; + is_uniform_ = 1; non_uniform_map_ = nullptr; - leaders_map_ = nullptr; - is_blocked_=0; + leaders_map_ = nullptr; + is_blocked_ = 0; } void Comm::destroy(Comm* comm) @@ -74,32 +51,44 @@ void Comm::destroy(Comm* comm) } int Comm::dup(MPI_Comm* newcomm){ - if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){ //we need to switch as the called function may silently touch global variables - smpi_switch_data_segment(smpi_process()->index()); - } + if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) { + // we need to switch as the called function may silently touch global variables + smpi_switch_data_segment(s4u::Actor::self()); + } MPI_Group cp = new Group(this->group()); - (*newcomm) = new Comm(cp, this->topo()); - int ret = MPI_SUCCESS; + (*newcomm) = new Comm(cp, this->topo()); + int ret = MPI_SUCCESS; if (not attributes()->empty()) { - int flag; - void* value_out; - for(auto it : *attributes()){ + 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}); } } - } } + } return ret; } @@ -110,10 +99,6 @@ MPI_Group Comm::group() return group_; } -MPI_Topology Comm::topo() { - return topo_; -} - int Comm::size() { if (this == MPI_COMM_UNINITIALIZED) @@ -125,7 +110,7 @@ int Comm::rank() { if (this == MPI_COMM_UNINITIALIZED) return smpi_process()->comm_world()->rank(); - return group_->rank(smpi_process()->index()); + return group_->rank(s4u::Actor::self()); } void Comm::get_name (char* name, int* len) @@ -134,14 +119,27 @@ 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",5); - *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 (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); @@ -150,10 +148,6 @@ 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(); @@ -190,6 +184,12 @@ 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) @@ -200,13 +200,13 @@ MPI_Comm Comm::split(int color, int key) 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; @@ -214,40 +214,37 @@ MPI_Comm Comm::split(int color, int key) Coll_gather_default::gather(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); - int* rankmap = xbt_new(int, 2 * size); + std::vector> rankmap; + rankmap.reserve(size); for (int i = 0; i < size; i++) { if (recvbuf[2 * i] != MPI_UNDEFINED) { - int count = 0; + rankmap.clear(); for (int j = i + 1; j < size; j++) { if(recvbuf[2 * i] == recvbuf[2 * j]) { recvbuf[2 * j] = MPI_UNDEFINED; - rankmap[2 * count] = j; - rankmap[2 * count + 1] = recvbuf[2 * j + 1]; - count++; + rankmap.push_back({recvbuf[2 * j + 1], j}); } } /* Add self in the group */ recvbuf[2 * i] = MPI_UNDEFINED; - rankmap[2 * count] = i; - rankmap[2 * count + 1] = recvbuf[2 * i + 1]; - count++; - qsort(rankmap, count, 2 * sizeof(int), &smpi_compare_rankmap); - group_out = new Group(count); + rankmap.push_back({recvbuf[2 * i + 1], i}); + std::sort(begin(rankmap), end(rankmap)); + group_out = new Group(rankmap.size()); if (i == 0) { group_root = group_out; /* Save root's group */ } - for (int j = 0; j < count; j++) { - int index = group->index(rankmap[2 * j]); - group_out->set_mapping(index, j); + for (unsigned j = 0; j < rankmap.size(); j++) { + s4u::ActorPtr actor = group->actor(rankmap[j].second); + group_out->set_mapping(actor, j); } - MPI_Request* requests = xbt_new(MPI_Request, count); + MPI_Request* requests = xbt_new(MPI_Request, rankmap.size()); int reqs = 0; - for (int j = 0; j < count; j++) { - if(rankmap[2 * j] != 0) { + for (auto const& rank : rankmap) { + if (rank.second != 0) { group_snd[reqs]=new Group(group_out); - requests[reqs] = Request::isend(&(group_snd[reqs]), 1, MPI_PTR, rankmap[2 * j], system_tag, this); + requests[reqs] = Request::isend(&(group_snd[reqs]), 1, MPI_PTR, rank.second, system_tag, this); reqs++; } } @@ -259,7 +256,6 @@ MPI_Comm Comm::split(int color, int key) } } xbt_free(recvbuf); - xbt_free(rankmap); xbt_free(group_snd); group_out = group_root; /* exit with root's group */ } else { @@ -284,10 +280,8 @@ void Comm::cleanup_smp(){ Comm::unref(intra_comm_); if (leaders_comm_ != MPI_COMM_NULL) Comm::unref(leaders_comm_); - if (non_uniform_map_ != nullptr) - xbt_free(non_uniform_map_); - if (leaders_map_ != nullptr) - xbt_free(leaders_map_); + xbt_free(non_uniform_map_); + delete[] leaders_map_; } void Comm::unref(Comm* comm){ @@ -317,75 +311,69 @@ void Comm::init_smp(){ // tell SimGrid we are not in replay for a while, because we need the buffers to be copied for the following calls bool replaying = false; //cache data to set it back again after if(smpi_process()->replaying()){ - replaying=true; - smpi_process()->set_replaying(false); + replaying = true; + smpi_process()->set_replaying(false); } - if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){ //we need to switch as the called function may silently touch global variables - smpi_switch_data_segment(smpi_process()->index()); - } + if (smpi_privatize_global_variables == 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 indexes of all processes sharing the same simix host - xbt_swag_t process_list = sg_host_self()->extension()->process_list; - int intra_comm_size = 0; - int min_index = INT_MAX; // the minimum index will be the leader - smx_actor_t actor = nullptr; - xbt_swag_foreach(actor, process_list) - { - int index = actor->pid - 1; - - if (this->group()->rank(index) != MPI_UNDEFINED) { - intra_comm_size++; - // the process is in the comm - if (index < min_index) - min_index = index; - } + //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; - actor = nullptr; - xbt_swag_foreach(actor, process_list) { - int index = actor->pid -1; - if(this->group()->rank(index)!=MPI_UNDEFINED){ - group_intra->set_mapping(index, i); + for (auto& actor : process_list) { + if (this->group()->rank(actor.iface()) != MPI_UNDEFINED) { + group_intra->set_mapping(actor.iface(), i); i++; } } - MPI_Comm comm_intra = new Comm(group_intra, nullptr); + MPI_Comm comm_intra = new Comm(group_intra, nullptr, 1); leader=min_index; - int * leaders_map= static_cast(xbt_malloc0(sizeof(int)*comm_size)); - int * leader_list= static_cast(xbt_malloc0(sizeof(int)*comm_size)); - for(i=0; iindex()); - } + if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) { + // we need to switch as the called function may silently touch global variables + smpi_switch_data_segment(s4u::Actor::self()); + } if(leaders_map_==nullptr){ leaders_map_= leaders_map; }else{ - xbt_free(leaders_map); + delete[] leaders_map; } - int j=0; int leader_group_size = 0; for(i=0; iset_mapping(leader_list[i], i); - leader_comm = new Comm(leaders_group, nullptr); + leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]), i); + leader_comm = new Comm(leaders_group, nullptr,1); this->set_leaders_comm(leader_comm); this->set_intra_comm(comm_intra); - //create intracommunicator + // create intracommunicator }else{ for (i=0; i< leader_group_size;i++) - leaders_group->set_mapping(leader_list[i], i); + leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]), 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(); @@ -420,7 +408,7 @@ 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, + Coll_allgather_ring::allgather(&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]) { @@ -429,27 +417,28 @@ void Comm::init_smp(){ } } if(is_uniform==0 && this->is_uniform()!=0){ - non_uniform_map_= non_uniform_map; + non_uniform_map_ = non_uniform_map; }else{ - xbt_free(non_uniform_map); + xbt_free(non_uniform_map); } is_uniform_=is_uniform; } - Coll_bcast_mpich::bcast(&(is_uniform_),1, MPI_INT, 0, comm_intra ); + Coll_bcast_scatter_LR_allgather::bcast(&(is_uniform_),1, MPI_INT, 0, comm_intra ); - if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){ //we need to switch as the called function may silently touch global variables - smpi_switch_data_segment(smpi_process()->index()); - } + if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) { + // we need to switch as the called function may silently touch global variables + smpi_switch_data_segment(s4u::Actor::self()); + } // Are the ranks blocked ? = allocated contiguously on the SMP nodes int is_blocked=1; - int prev=this->group()->rank(comm_intra->group()->index(0)); - for (i=1; igroup()->rank(comm_intra->group()->index(i)); - if(that!=prev+1){ - is_blocked=0; - break; - } - prev = that; + 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) { + is_blocked = 0; + break; + } + prev = that; } int global_blocked; @@ -457,12 +446,12 @@ void Comm::init_smp(){ if(MPI_COMM_WORLD==MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD){ if(this->rank()==0){ - is_blocked_=global_blocked; + is_blocked_ = global_blocked; } }else{ is_blocked_=global_blocked; } - xbt_free(leader_list); + delete[] leader_list; if(replaying) smpi_process()->set_replaying(true); @@ -474,9 +463,10 @@ MPI_Comm Comm::f2c(int id) { } else if(id==0){ return MPI_COMM_WORLD; } else if(F2C::f2c_lookup() != nullptr && id >= 0) { - char key[KEY_SIZE]; - MPI_Comm tmp = static_cast(xbt_dict_get_or_null(F2C::f2c_lookup(),get_key_id(key, id))); - return tmp != nullptr ? tmp : MPI_COMM_NULL ; + char key[KEY_SIZE]; + const auto& lookup = F2C::f2c_lookup(); + auto comm = lookup->find(get_key_id(key, id)); + return comm == lookup->end() ? MPI_COMM_NULL : static_cast(comm->second); } else { return MPI_COMM_NULL; } @@ -484,20 +474,19 @@ MPI_Comm Comm::f2c(int id) { void Comm::free_f(int id) { char key[KEY_SIZE]; - xbt_dict_remove(F2C::f2c_lookup(), id==0? get_key(key, id) : get_key_id(key, id)); + 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(xbt_dict_new_homogeneous(nullptr)); + F2C::set_f2c_lookup(new std::unordered_map); } char key[KEY_SIZE]; - xbt_dict_set(F2C::f2c_lookup(), this==MPI_COMM_WORLD? get_key(key, F2C::f2c_id()) : get_key_id(key,F2C::f2c_id()), this, nullptr); + (*(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; } - void Comm::add_rma_win(MPI_Win win){ rma_wins_.push_back(win); } @@ -507,7 +496,7 @@ void Comm::remove_rma_win(MPI_Win win){ } void Comm::finish_rma_calls(){ - for(auto it : rma_wins_){ + for (auto const& it : rma_wins_) { if(it->rank()==this->rank()){//is it ours (for MPI_COMM_WORLD)? int finished = it->finish_comms(); XBT_DEBUG("Barrier for rank %d - Finished %d RMA calls",this->rank(), finished); @@ -515,8 +504,29 @@ void Comm::finish_rma_calls(){ } } - +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_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(); +} +} // namespace smpi +} // namespace simgrid