X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/d656f9465e7b1c36319f86eeafdcec58fe9551a4..2d37e348a09783cda723c7019640ee69de168324:/src/smpi/mpi/smpi_win.cpp diff --git a/src/smpi/mpi/smpi_win.cpp b/src/smpi/mpi/smpi_win.cpp index 7cc829aa18..cefc4c567a 100644 --- a/src/smpi/mpi/smpi_win.cpp +++ b/src/smpi/mpi/smpi_win.cpp @@ -1,20 +1,23 @@ -/* Copyright (c) 2007-2017. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2019. 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_win.hpp" + #include "private.hpp" #include "smpi_coll.hpp" #include "smpi_comm.hpp" #include "smpi_datatype.hpp" #include "smpi_info.hpp" #include "smpi_keyvals.hpp" -#include "smpi_process.hpp" #include "smpi_request.hpp" +#include "src/smpi/include/smpi_actor.hpp" XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_rma, smpi, "Logging specific to SMPI (RMA operations)"); +using simgrid::s4u::Actor; + namespace simgrid{ namespace smpi{ std::unordered_map Win::keyvals_; @@ -22,38 +25,39 @@ int Win::keyval_id_=0; Win::Win(void *base, MPI_Aint size, int disp_unit, MPI_Info info, MPI_Comm comm, int allocated, int dynamic): base_(base), size_(size), disp_unit_(disp_unit), assert_(0), info_(info), comm_(comm), allocated_(allocated), dynamic_(dynamic){ int comm_size = comm->size(); - rank_ = comm->rank(); + rank_ = comm->rank(); XBT_DEBUG("Creating window"); if(info!=MPI_INFO_NULL) info->ref(); - name_ = nullptr; - opened_ = 0; - group_ = MPI_GROUP_NULL; - requests_ = new std::vector(); - mut_=xbt_mutex_init(); - lock_mut_=xbt_mutex_init(); - atomic_mut_=xbt_mutex_init(); - connected_wins_ = new MPI_Win[comm_size]; + name_ = nullptr; + opened_ = 0; + group_ = MPI_GROUP_NULL; + requests_ = new std::vector(); + mut_ = xbt_mutex_init(); + lock_mut_ = xbt_mutex_init(); + atomic_mut_ = xbt_mutex_init(); + connected_wins_ = new MPI_Win[comm_size]; connected_wins_[rank_] = this; - count_ = 0; + count_ = 0; if(rank_==0){ - bar_ = MSG_barrier_init(comm_size); + bar_ = new simgrid::s4u::Barrier(comm_size); } mode_=0; comm->add_rma_win(this); + comm->ref(); Colls::allgather(&(connected_wins_[rank_]), sizeof(MPI_Win), MPI_BYTE, connected_wins_, sizeof(MPI_Win), MPI_BYTE, comm); - Colls::bcast(&(bar_), sizeof(msg_bar_t), MPI_BYTE, 0, comm); + Colls::bcast(&(bar_), sizeof(simgrid::s4u::Barrier*), MPI_BYTE, 0, comm); Colls::barrier(comm); } Win::~Win(){ //As per the standard, perform a barrier to ensure every async comm is finished - MSG_barrier_wait(bar_); + bar_->wait(); int finished = finish_comms(); XBT_DEBUG("Win destructor - Finished %d RMA calls", finished); @@ -70,9 +74,10 @@ Win::~Win(){ comm_->remove_rma_win(this); Colls::barrier(comm_); - int rank=comm_->rank(); - if(rank == 0) - MSG_barrier_destroy(bar_); + Comm::unref(comm_); + + if (rank_ == 0) + delete bar_; xbt_mutex_destroy(mut_); xbt_mutex_destroy(lock_mut_); xbt_mutex_destroy(atomic_mut_); @@ -159,7 +164,7 @@ int Win::fence(int assert) opened_=1; if (assert != MPI_MODE_NOPRECEDE) { // This is not the first fence => finalize what came before - MSG_barrier_wait(bar_); + bar_->wait(); xbt_mutex_acquire(mut_); // This (simulated) mutex ensures that no process pushes to the vector of requests during the waitall. // Without this, the vector could get redimensionned when another process pushes. @@ -180,7 +185,7 @@ int Win::fence(int assert) opened_=0; assert_ = assert; - MSG_barrier_wait(bar_); + bar_->wait(); XBT_DEBUG("Leaving fence"); return MPI_SUCCESS; @@ -206,16 +211,19 @@ int Win::put( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, return MPI_ERR_ARG; void* recv_addr = static_cast ( static_cast(recv_win->base_) + target_disp * recv_win->disp_unit_); - XBT_DEBUG("Entering MPI_Put to %d", target_rank); - if(target_rank != comm_->rank()){ - //prepare send_request - MPI_Request sreq = Request::rma_send_init(origin_addr, origin_count, origin_datatype, smpi_process()->index(), - comm_->group()->index(target_rank), SMPI_RMA_TAG+1, comm_, MPI_OP_NULL); + if (target_rank != comm_->rank()) { // This is not for myself, so we need to send messages + XBT_DEBUG("Entering MPI_Put to remote rank %d", target_rank); + // prepare send_request + MPI_Request sreq = + // TODO cheinrich Check for rank / pid conversion + Request::rma_send_init(origin_addr, origin_count, origin_datatype, comm_->rank(), target_rank, SMPI_RMA_TAG + 1, + comm_, MPI_OP_NULL); //prepare receiver request - MPI_Request rreq = Request::rma_recv_init(recv_addr, target_count, target_datatype, smpi_process()->index(), - comm_->group()->index(target_rank), SMPI_RMA_TAG+1, recv_win->comm_, MPI_OP_NULL); + // TODO cheinrich Check for rank / pid conversion + MPI_Request rreq = Request::rma_recv_init(recv_addr, target_count, target_datatype, recv_win->comm_->rank(), + target_rank, SMPI_RMA_TAG + 1, recv_win->comm_, MPI_OP_NULL); //start send sreq->start(); @@ -235,6 +243,7 @@ int Win::put( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, xbt_mutex_release(recv_win->mut_); }else{ + XBT_DEBUG("Entering MPI_Put from myself to myself, rank %d", target_rank); Datatype::copy(origin_addr, origin_count, origin_datatype, recv_addr, target_count, target_datatype); if(request!=nullptr) *request = MPI_REQUEST_NULL; @@ -267,14 +276,14 @@ int Win::get( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, if(target_rank != comm_->rank()){ //prepare send_request - MPI_Request sreq = Request::rma_send_init(send_addr, target_count, target_datatype, - comm_->group()->index(target_rank), smpi_process()->index(), SMPI_RMA_TAG+2, send_win->comm_, - MPI_OP_NULL); + MPI_Request sreq = Request::rma_send_init(send_addr, target_count, target_datatype, target_rank, + send_win->comm_->rank(), SMPI_RMA_TAG + 2, send_win->comm_, MPI_OP_NULL); //prepare receiver request - MPI_Request rreq = Request::rma_recv_init(origin_addr, origin_count, origin_datatype, - comm_->group()->index(target_rank), smpi_process()->index(), SMPI_RMA_TAG+2, comm_, - MPI_OP_NULL); + MPI_Request rreq = Request::rma_recv_init( + origin_addr, origin_count, origin_datatype, target_rank, + comm_->rank(), // TODO cheinrich Check here if comm_->rank() and above send_win->comm_->rank() are correct + SMPI_RMA_TAG + 2, comm_, MPI_OP_NULL); //start the send, with another process than us as sender. sreq->start(); @@ -307,7 +316,7 @@ int Win::get( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, int Win::accumulate( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, int target_rank, MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype, MPI_Op op, MPI_Request* request) { - + XBT_DEBUG("Entering MPI_Win_Accumulate"); //get receiver pointer MPI_Win recv_win = connected_wins_[target_rank]; @@ -330,31 +339,32 @@ int Win::accumulate( void *origin_addr, int origin_count, MPI_Datatype origin_da //As the tag will be used for ordering of the operations, substract count from it (to avoid collisions with other SMPI tags, SMPI_RMA_TAG is set below all the other ones we use ) //prepare send_request - MPI_Request sreq = Request::rma_send_init(origin_addr, origin_count, origin_datatype, - smpi_process()->index(), comm_->group()->index(target_rank), SMPI_RMA_TAG-3-count_, comm_, op); + MPI_Request sreq = Request::rma_send_init(origin_addr, origin_count, origin_datatype, comm_->rank(), target_rank, + SMPI_RMA_TAG - 3 - count_, comm_, op); - //prepare receiver request - MPI_Request rreq = Request::rma_recv_init(recv_addr, target_count, target_datatype, - smpi_process()->index(), comm_->group()->index(target_rank), SMPI_RMA_TAG-3-count_, recv_win->comm_, op); + // prepare receiver request + MPI_Request rreq = Request::rma_recv_init(recv_addr, target_count, target_datatype, recv_win->comm_->rank(), + recv_win->comm_->group()->rank(comm_->group()->actor(target_rank)), SMPI_RMA_TAG - 3 - count_, recv_win->comm_, op); - count_++; + count_++; - //start send - sreq->start(); - //push request to receiver's win - xbt_mutex_acquire(recv_win->mut_); - recv_win->requests_->push_back(rreq); - rreq->start(); - xbt_mutex_release(recv_win->mut_); + // start send + sreq->start(); + // push request to receiver's win + xbt_mutex_acquire(recv_win->mut_); + recv_win->requests_->push_back(rreq); + rreq->start(); + xbt_mutex_release(recv_win->mut_); - if(request!=nullptr){ - *request=sreq; - }else{ - xbt_mutex_acquire(mut_); - requests_->push_back(sreq); - xbt_mutex_release(mut_); - } + if (request != nullptr) { + *request = sreq; + } else { + xbt_mutex_acquire(mut_); + requests_->push_back(sreq); + xbt_mutex_release(mut_); + } + XBT_DEBUG("Leaving MPI_Win_Accumulate"); return MPI_SUCCESS; } @@ -428,42 +438,44 @@ int Win::compare_and_swap(void *origin_addr, void *compare_addr, } int Win::start(MPI_Group group, int assert){ - /* From MPI forum advices - The call to MPI_WIN_COMPLETE does not return until the put call has completed at the origin; and the target window - will be accessed by the put operation only after the call to MPI_WIN_START has matched a call to MPI_WIN_POST by - the target process. This still leaves much choice to implementors. The call to MPI_WIN_START can block until the - matching call to MPI_WIN_POST occurs at all target processes. One can also have implementations where the call to - MPI_WIN_START is nonblocking, but the call to MPI_PUT blocks until the matching call to MPI_WIN_POST occurred; or - implementations where the first two calls are nonblocking, but the call to MPI_WIN_COMPLETE blocks until the call - to MPI_WIN_POST occurred; or even implementations where all three calls can complete before any target process - called MPI_WIN_POST --- the data put must be buffered, in this last case, so as to allow the put to complete at the - origin ahead of its completion at the target. However, once the call to MPI_WIN_POST is issued, the sequence above - must complete, without further dependencies. */ + /* From MPI forum advices + The call to MPI_WIN_COMPLETE does not return until the put call has completed at the origin; and the target window + will be accessed by the put operation only after the call to MPI_WIN_START has matched a call to MPI_WIN_POST by + the target process. This still leaves much choice to implementors. The call to MPI_WIN_START can block until the + matching call to MPI_WIN_POST occurs at all target processes. One can also have implementations where the call to + MPI_WIN_START is nonblocking, but the call to MPI_PUT blocks until the matching call to MPI_WIN_POST occurred; or + implementations where the first two calls are nonblocking, but the call to MPI_WIN_COMPLETE blocks until the call + to MPI_WIN_POST occurred; or even implementations where all three calls can complete before any target process + called MPI_WIN_POST --- the data put must be buffered, in this last case, so as to allow the put to complete at the + origin ahead of its completion at the target. However, once the call to MPI_WIN_POST is issued, the sequence above + must complete, without further dependencies. */ //naive, blocking implementation. - int i = 0; - int j = 0; - int size = group->size(); - MPI_Request* reqs = xbt_new0(MPI_Request, size); - - while (j != size) { - int src = group->index(j); - if (src != smpi_process()->index() && src != MPI_UNDEFINED) { - reqs[i] = Request::irecv_init(nullptr, 0, MPI_CHAR, src, SMPI_RMA_TAG + 4, MPI_COMM_WORLD); - i++; - } - j++; + int i = 0; + int j = 0; + int size = group->size(); + MPI_Request* reqs = xbt_new0(MPI_Request, size); + + XBT_DEBUG("Entering MPI_Win_Start"); + while (j != size) { + int src = comm_->group()->rank(group->actor(j)); + if (src != rank_ && src != MPI_UNDEFINED) { // TODO cheinrich: The check of MPI_UNDEFINED should be useless here + reqs[i] = Request::irecv_init(nullptr, 0, MPI_CHAR, src, SMPI_RMA_TAG + 4, comm_); + i++; + } + j++; } - size=i; + size = i; Request::startall(size, reqs); Request::waitall(size, reqs, MPI_STATUSES_IGNORE); - for(i=0;iref(); + XBT_DEBUG("Leaving MPI_Win_Start"); return MPI_SUCCESS; } @@ -474,10 +486,11 @@ int Win::post(MPI_Group group, int assert){ int size = group->size(); MPI_Request* reqs = xbt_new0(MPI_Request, size); + XBT_DEBUG("Entering MPI_Win_Post"); while(j!=size){ - int dst=group->index(j); - if(dst!=smpi_process()->index() && dst!=MPI_UNDEFINED){ - reqs[i]=Request::send_init(nullptr, 0, MPI_CHAR, dst, SMPI_RMA_TAG+4, MPI_COMM_WORLD); + int dst = comm_->group()->rank(group->actor(j)); + if (dst != rank_ && dst != MPI_UNDEFINED) { + reqs[i] = Request::send_init(nullptr, 0, MPI_CHAR, dst, SMPI_RMA_TAG + 4, comm_); i++; } j++; @@ -493,6 +506,7 @@ int Win::post(MPI_Group group, int assert){ opened_++; //we're open for business ! group_=group; group->ref(); + XBT_DEBUG("Leaving MPI_Win_Post"); return MPI_SUCCESS; } @@ -503,13 +517,13 @@ int Win::complete(){ XBT_DEBUG("Entering MPI_Win_Complete"); int i = 0; int j = 0; - int size = group_->size(); + int size = group_->size(); MPI_Request* reqs = xbt_new0(MPI_Request, size); while(j!=size){ - int dst=group_->index(j); - if(dst!=smpi_process()->index() && dst!=MPI_UNDEFINED){ - reqs[i]=Request::send_init(nullptr, 0, MPI_CHAR, dst, SMPI_RMA_TAG+5, MPI_COMM_WORLD); + int dst = comm_->group()->rank(group_->actor(j)); + if (dst != rank_ && dst != MPI_UNDEFINED) { + reqs[i] = Request::send_init(nullptr, 0, MPI_CHAR, dst, SMPI_RMA_TAG + 5, comm_); i++; } j++; @@ -541,9 +555,9 @@ int Win::wait(){ MPI_Request* reqs = xbt_new0(MPI_Request, size); while(j!=size){ - int src=group_->index(j); - if(src!=smpi_process()->index() && src!=MPI_UNDEFINED){ - reqs[i]=Request::irecv_init(nullptr, 0, MPI_CHAR, src,SMPI_RMA_TAG+5, MPI_COMM_WORLD); + int src = comm_->group()->rank(group_->actor(j)); + if (src != rank_ && src != MPI_UNDEFINED) { + reqs[i] = Request::irecv_init(nullptr, 0, MPI_CHAR, src, SMPI_RMA_TAG + 5, comm_); i++; } j++; @@ -616,18 +630,18 @@ int Win::unlock_all(){ int i=0; int retval = MPI_SUCCESS; for (i=0; isize();i++){ - int ret = this->unlock(i); - if(ret != MPI_SUCCESS) - retval = ret; + int ret = this->unlock(i); + if (ret != MPI_SUCCESS) + retval = ret; } return retval; } int Win::flush(int rank){ MPI_Win target_win = connected_wins_[rank]; - int finished = finish_comms(rank); + int finished = finish_comms(rank_); XBT_DEBUG("Win_flush on local %d - Finished %d RMA calls", rank_, finished); - finished = target_win->finish_comms(rank_); + finished = target_win->finish_comms(rank); XBT_DEBUG("Win_flush on remote %d - Finished %d RMA calls", rank, finished); return MPI_SUCCESS; } @@ -639,11 +653,9 @@ int Win::flush_local(int rank){ } int Win::flush_all(){ - int i=0; - int finished = 0; - finished = finish_comms(); + int finished = finish_comms(); XBT_DEBUG("Win_flush_all on local - Finished %d RMA calls", finished); - for (i=0; isize();i++){ + for (int i = 0; i < comm_->size(); i++) { finished = connected_wins_[i]->finish_comms(rank_); XBT_DEBUG("Win_flush_all on %d - Finished %d RMA calls", i, finished); } @@ -660,7 +672,6 @@ Win* Win::f2c(int id){ return static_cast(F2C::f2c(id)); } - int Win::finish_comms(){ xbt_mutex_acquire(mut_); //Finish own requests @@ -684,8 +695,13 @@ int Win::finish_comms(int rank){ size = 0; std::vector myreqqs; std::vector::iterator iter = reqqs->begin(); + int proc_id = comm_->group()->actor(rank)->get_pid(); while (iter != reqqs->end()){ - if(((*iter)!=MPI_REQUEST_NULL) && (((*iter)->src() == rank) || ((*iter)->dst() == rank))){ + // Let's see if we're either the destination or the sender of this request + // because we only wait for requests that we are responsible for. + // Also use the process id here since the request itself returns from src() + // and dst() the process id, NOT the rank (which only exists in the context of a communicator). + if (((*iter) != MPI_REQUEST_NULL) && (((*iter)->src() == proc_id) || ((*iter)->dst() == proc_id))) { myreqqs.push_back(*iter); iter = reqqs->erase(iter); size++; @@ -703,6 +719,22 @@ int Win::finish_comms(int rank){ return size; } - +int Win::shared_query(int rank, MPI_Aint* size, int* disp_unit, void* baseptr) +{ + MPI_Win target_win = rank != MPI_PROC_NULL ? connected_wins_[rank] : nullptr; + for (int i = 0; not target_win && i < comm_->size(); i++) { + if (connected_wins_[i]->size_ > 0) + target_win = connected_wins_[i]; + } + if (target_win) { + *size = target_win->size_; + *disp_unit = target_win->disp_unit_; + *static_cast(baseptr) = target_win->base_; + } else { + *size = 0; + *static_cast(baseptr) = xbt_malloc(0); + } + return MPI_SUCCESS; +} } }