X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/7a1d0ed1e978b9c32b1e0a73011f76c3b9e7033f..a66cfa3e250bfd537fb32d32f71c44115beeb6f6:/src/smpi/mpi/smpi_request.cpp diff --git a/src/smpi/mpi/smpi_request.cpp b/src/smpi/mpi/smpi_request.cpp index d34a031538..58892ed47c 100644 --- a/src/smpi/mpi/smpi_request.cpp +++ b/src/smpi/mpi/smpi_request.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2007-2019. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2020. 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. */ @@ -33,26 +33,18 @@ extern void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl* namespace simgrid{ namespace smpi{ -Request::Request(const void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags, MPI_Op op) - : buf_(const_cast(buf)), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags), op_(op) +Request::Request(const void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, + unsigned flags, MPI_Op op) + : buf_(const_cast(buf)) + , old_type_(datatype) + , size_(datatype->size() * count) + , src_(src) + , dst_(dst) + , tag_(tag) + , comm_(comm) + , flags_(flags) + , op_(op) { - void *old_buf = nullptr; -// FIXME Handle the case of a partial shared malloc. - if ((((flags & MPI_REQ_RECV) != 0) && ((flags & MPI_REQ_ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) { - // This part handles the problem of non-contiguous memory - old_buf = const_cast(buf); - if (count==0){ - buf_ = nullptr; - }else { - buf_ = xbt_malloc(count*datatype->size()); - if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & MPI_REQ_SEND) != 0)) { - datatype->serialize(old_buf, buf_, count); - } - } - } - // This part handles the problem of non-contiguous memory (for the unserialization at the reception) - old_buf_ = old_buf; - size_ = datatype->size() * count; datatype->ref(); comm_->ref(); if(op != MPI_REPLACE && op != MPI_OP_NULL) @@ -72,6 +64,7 @@ Request::Request(const void* buf, int count, MPI_Datatype datatype, int src, int generalized_funcs=nullptr; nbc_requests_=nullptr; nbc_requests_size_=0; + init_buffer(count); } void Request::ref(){ @@ -108,58 +101,68 @@ void Request::unref(MPI_Request* request) } } -int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*) +bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request receiver) { - MPI_Request ref = static_cast(a); - MPI_Request req = static_cast(b); - XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d, id %d against %d",ref->src_,req->src_, ref->tag_, req->tag_,ref->comm_->id(),req->comm_->id()); - - xbt_assert(ref, "Cannot match recv against null reference"); - xbt_assert(req, "Cannot match recv against null request"); - if((ref->comm_->id()==MPI_UNDEFINED || req->comm_->id() == MPI_UNDEFINED || (ref->comm_->id()==req->comm_->id())) - && ((ref->src_ == MPI_ANY_SOURCE && (ref->comm_->group()->rank(req->src_) != MPI_UNDEFINED)) || req->src_ == ref->src_) - && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){ - //we match, we can transfer some values - if(ref->src_ == MPI_ANY_SOURCE) - ref->real_src_ = req->src_; - if(ref->tag_ == MPI_ANY_TAG) - ref->real_tag_ = req->tag_; - if(ref->real_size_ < req->real_size_) - ref->truncated_ = true; - if (req->detached_) - ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver - if(req->cancelled_==0) + xbt_assert(sender, "Cannot match against null sender"); + xbt_assert(receiver, "Cannot match against null receiver"); + XBT_DEBUG("Trying to match %s of sender src %d against %d, tag %d against %d, id %d against %d", + (req == receiver ? "send" : "recv"), sender->src_, receiver->src_, sender->tag_, receiver->tag_, + sender->comm_->id(), receiver->comm_->id()); + + if ((receiver->comm_->id() == MPI_UNDEFINED || sender->comm_->id() == MPI_UNDEFINED || + receiver->comm_->id() == sender->comm_->id()) && + ((receiver->src_ == MPI_ANY_SOURCE && (receiver->comm_->group()->rank(sender->src_) != MPI_UNDEFINED)) || + receiver->src_ == sender->src_) && + ((receiver->tag_ == MPI_ANY_TAG && sender->tag_ >= 0) || receiver->tag_ == sender->tag_)) { + // we match, we can transfer some values + if (receiver->src_ == MPI_ANY_SOURCE) + receiver->real_src_ = sender->src_; + if (receiver->tag_ == MPI_ANY_TAG) + receiver->real_tag_ = sender->tag_; + if (receiver->real_size_ < sender->real_size_) + receiver->truncated_ = true; + if (sender->detached_) + receiver->detached_sender_ = sender; // tie the sender to the receiver, as it is detached and has to be freed in + // the receiver + if (req->cancelled_ == 0) req->cancelled_ = -1; // mark as uncancelable XBT_DEBUG("match succeeded"); - return 1; - }else return 0; + return true; + } + return false; +} + +void Request::init_buffer(int count){ + void *old_buf = nullptr; +// FIXME Handle the case of a partial shared malloc. + // This part handles the problem of non-contiguous memory (for the unserialization at the reception) + if ((((flags_ & MPI_REQ_RECV) != 0) && ((flags_ & MPI_REQ_ACCUMULATE) != 0)) || (old_type_->flags() & DT_FLAG_DERIVED)) { + // This part handles the problem of non-contiguous memory + old_buf = const_cast(buf_); + if (count==0){ + buf_ = nullptr; + }else { + buf_ = xbt_malloc(count*old_type_->size()); + if ((old_type_->flags() & DT_FLAG_DERIVED) && ((flags_ & MPI_REQ_SEND) != 0)) { + old_type_->serialize(old_buf, buf_, count); + } + } + } + old_buf_ = old_buf; } -int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*) +bool Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*) { MPI_Request ref = static_cast(a); MPI_Request req = static_cast(b); - XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d, id %d against %d",ref->src_,req->src_, ref->tag_, req->tag_,ref->comm_->id(),req->comm_->id()); - xbt_assert(ref, "Cannot match send against null reference"); - xbt_assert(req, "Cannot match send against null request"); - - if((ref->comm_->id()==MPI_UNDEFINED || req->comm_->id() == MPI_UNDEFINED || (ref->comm_->id()==req->comm_->id())) - && ((req->src_ == MPI_ANY_SOURCE && (req->comm_->group()->rank(ref->src_) != MPI_UNDEFINED)) || req->src_ == ref->src_) - && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){ - if(req->src_ == MPI_ANY_SOURCE) - req->real_src_ = ref->src_; - if(req->tag_ == MPI_ANY_TAG) - req->real_tag_ = ref->tag_; - if(req->real_size_ < ref->real_size_) - req->truncated_ = true; - if (ref->detached_) - req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver - if(req->cancelled_==0) - req->cancelled_ = -1; // mark as uncancelable - XBT_DEBUG("match succeeded"); - return 1; - } else - return 0; + return match_common(req, req, ref); +} + +bool Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*) +{ + MPI_Request ref = static_cast(a); + MPI_Request req = static_cast(b); + return match_common(req, ref, req); } void Request::print_request(const char *message) @@ -168,11 +171,9 @@ void Request::print_request(const char *message) message, this, buf_, size_, src_, dst_, tag_, flags_); } - /* factories, to hide the internal flags from the caller */ MPI_Request Request::bsend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { - return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(), comm->group()->actor(dst)->get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED | MPI_REQ_BSEND); @@ -180,7 +181,6 @@ MPI_Request Request::bsend_init(const void *buf, int count, MPI_Datatype datatyp MPI_Request Request::send_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { - return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(), comm->group()->actor(dst)->get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED); @@ -367,6 +367,11 @@ void Request::start() s4u::Mailbox* mailbox; xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication"); + //reinitialize temporary buffer for persistent requests + if(real_size_ > 0 && flags_ & MPI_REQ_FINISHED){ + buf_ = old_buf_; + init_buffer(real_size_/old_type_->size()); + } flags_ &= ~MPI_REQ_PREPARED; flags_ &= ~MPI_REQ_FINISHED; this->ref(); @@ -376,21 +381,19 @@ void Request::start() simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_)); - int async_small_thresh = simgrid::config::get_value("smpi/async-small-thresh"); - simgrid::s4u::MutexPtr mut = process->mailboxes_mutex(); - if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0) + if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0) mut->lock(); - if (async_small_thresh == 0 && (flags_ & MPI_REQ_RMA) == 0) { + if (smpi_cfg_async_small_thresh() == 0 && (flags_ & MPI_REQ_RMA) == 0) { mailbox = process->mailbox(); - } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast(size_) < async_small_thresh) { + } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast(size_) < smpi_cfg_async_small_thresh()) { //We have to check both mailboxes (because SSEND messages are sent to the large mbox). //begin with the more appropriate one : the small one. mailbox = process->mailbox_small(); XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %s (in case of SSEND)?", mailbox->get_cname()); - smx_activity_t action = mailbox->iprobe(0, &match_recv, static_cast(this)); + simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast(this)); if (action == nullptr) { mailbox = process->mailbox(); @@ -406,7 +409,7 @@ void Request::start() } else { mailbox = process->mailbox_small(); XBT_DEBUG("Is there a corresponding send already posted the small mailbox?"); - smx_activity_t action = mailbox->iprobe(0, &match_recv, static_cast(this)); + simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast(this)); if (action == nullptr) { XBT_DEBUG("No, nothing in the permanent receive mailbox"); @@ -423,7 +426,7 @@ void Request::start() process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0); XBT_DEBUG("recv simcall posted"); - if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0) + if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0) mut->unlock(); } else { /* the RECV flag was not set, so this is a send */ simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_)); @@ -437,7 +440,7 @@ void Request::start() void* buf = buf_; if ((flags_ & MPI_REQ_SSEND) == 0 && ((flags_ & MPI_REQ_RMA) != 0 || (flags_ & MPI_REQ_BSEND) != 0 || - static_cast(size_) < simgrid::config::get_value("smpi/send-is-detached-thresh"))) { + static_cast(size_) < smpi_cfg_detached_send_thresh())) { void *oldbuf = nullptr; detached_ = true; XBT_DEBUG("Send request %p is detached", this); @@ -445,7 +448,7 @@ void Request::start() if (not(old_type_->flags() & DT_FLAG_DERIVED)) { oldbuf = buf_; if (not process->replaying() && oldbuf != nullptr && size_ != 0) { - if ((smpi_privatize_global_variables != SmpiPrivStrategies::NONE) && + if ((smpi_cfg_privatization() != SmpiPrivStrategies::NONE) && (static_cast(buf_) >= smpi_data_exe_start) && (static_cast(buf_) < smpi_data_exe_start + smpi_data_exe_size)) { XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment "); @@ -474,19 +477,17 @@ void Request::start() XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime); } - int async_small_thresh = simgrid::config::get_value("smpi/async-small-thresh"); - simgrid::s4u::MutexPtr mut = process->mailboxes_mutex(); - if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0) + if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0) mut->lock(); - if (not(async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)) { + if (not(smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)) { mailbox = process->mailbox(); - } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast(size_) < async_small_thresh) { // eager mode + } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast(size_) < smpi_cfg_async_small_thresh()) { // eager mode mailbox = process->mailbox(); XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %s?", mailbox->get_cname()); - smx_activity_t action = mailbox->iprobe(1, &match_send, static_cast(this)); + simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(1, &match_send, static_cast(this)); if (action == nullptr) { if ((flags_ & MPI_REQ_SSEND) == 0) { mailbox = process->mailbox_small(); @@ -512,8 +513,9 @@ void Request::start() // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later real_size_=size_; + size_t payload_size_ = size_ + 16;//MPI enveloppe size (tag+dest+communicator) action_ = simcall_comm_isend( - simgrid::s4u::Actor::by_pid(src_)->get_impl(), mailbox->get_impl(), size_, -1.0, buf, real_size_, &match_send, + simgrid::s4u::Actor::by_pid(src_)->get_impl(), mailbox->get_impl(), payload_size_, -1.0, buf, real_size_, &match_send, &xbt_free_f, // how to free the userdata if a detached send fails process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, // detach if msg size < eager/rdv switch limit @@ -526,7 +528,7 @@ void Request::start() smpi_process()->get_tracing_category()); } - if (async_small_thresh != 0 || ((flags_ & MPI_REQ_RMA) != 0)) + if (smpi_cfg_async_small_thresh() != 0 || ((flags_ & MPI_REQ_RMA) != 0)) mut->unlock(); } } @@ -575,10 +577,10 @@ int Request::test(MPI_Request * request, MPI_Status * status, int* flag) { Status::empty(status); *flag = 1; - if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) { + if (((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) == 0) { if ((*request)->action_ != nullptr && (*request)->cancelled_ != 1){ try{ - *flag = simcall_comm_test((*request)->action_); + *flag = simcall_comm_test((*request)->action_.get()); } catch (const Exception&) { *flag = 0; return ret; @@ -605,7 +607,7 @@ int Request::test(MPI_Request * request, MPI_Status * status, int* flag) { nsleeps=1;//reset the number of sleeps we will do next time if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0) *request = MPI_REQUEST_NULL; - } else if (simgrid::config::get_value("smpi/grow-injected-times")) { + } else if (smpi_cfg_grow_injected_times()) { nsleeps++; } } @@ -763,7 +765,7 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* // This can speed up the execution of certain applications by an order of magnitude, such as HPL static int nsleeps = 1; double speed = s4u::this_actor::get_host()->get_speed(); - double maxrate = simgrid::config::get_value("smpi/iprobe-cpu-usage"); + double maxrate = smpi_cfg_iprobe_cpu_usage(); MPI_Request request = new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source)->get_pid(), simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV); @@ -786,7 +788,7 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* request->print_request("New iprobe"); // We have to test both mailboxes as we don't know if we will receive one or another - if (simgrid::config::get_value("smpi/async-small-thresh") > 0) { + if (smpi_cfg_async_small_thresh() > 0) { mailbox = smpi_process()->mailbox_small(); XBT_DEBUG("Trying to probe the perm recv mailbox"); request->action_ = mailbox->iprobe(0, &match_recv, static_cast(request)); @@ -800,7 +802,7 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* if (request->action_ != nullptr){ kernel::activity::CommImplPtr sync_comm = boost::static_pointer_cast(request->action_); - MPI_Request req = static_cast(sync_comm->src_data_); + const Request* req = static_cast(sync_comm->src_data_); *flag = 1; if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) { status->MPI_SOURCE = comm->group()->rank(req->src_); @@ -812,7 +814,7 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* } else { *flag = 0; - if (simgrid::config::get_value("smpi/grow-injected-times")) + if (smpi_cfg_grow_injected_times()) nsleeps++; } unref(&request); @@ -850,8 +852,7 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status) // FIXME Handle the case of a partial shared malloc. if (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){ - - if (not smpi_process()->replaying() && smpi_privatize_global_variables != SmpiPrivStrategies::NONE && + if (not smpi_process()->replaying() && smpi_cfg_privatization() != SmpiPrivStrategies::NONE && static_cast(req->old_buf_) >= smpi_data_exe_start && static_cast(req->old_buf_) < smpi_data_exe_start + smpi_data_exe_size) { XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment "); @@ -863,12 +864,14 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status) if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0) datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_); xbt_free(req->buf_); + req->buf_=nullptr; } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate if (datatype->size() != 0) { int n = req->real_size_ / datatype->size(); req->op_->apply(req->buf_, req->old_buf_, &n, datatype); } xbt_free(req->buf_); + req->buf_=nullptr; } } } @@ -926,7 +929,7 @@ int Request::wait(MPI_Request * request, MPI_Status * status) } (*request)->print_request("Waiting"); - if ((*request)->flags_ & MPI_REQ_PREPARED) { + if ((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) { Status::empty(status); return ret; } @@ -934,7 +937,7 @@ int Request::wait(MPI_Request * request, MPI_Status * status) if ((*request)->action_ != nullptr){ try{ // this is not a detached send - simcall_comm_wait((*request)->action_, -1.0); + simcall_comm_wait((*request)->action_.get(), -1.0); } catch (const Exception&) { XBT_VERB("Request cancelled"); } @@ -1112,7 +1115,8 @@ int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Sta return count; } -MPI_Request Request::f2c(int id) { +MPI_Request Request::f2c(int id) +{ char key[KEY_SIZE]; if(id==MPI_FORTRAN_REQUEST_NULL) return static_cast(MPI_REQUEST_NULL); @@ -1127,7 +1131,8 @@ void Request::free_f(int id) } } -int Request::get_status(MPI_Request req, int* flag, MPI_Status * status){ +int Request::get_status(const Request* req, int* flag, MPI_Status* status) +{ *flag=0; if(req != MPI_REQUEST_NULL && req->action_ != nullptr) { @@ -1154,8 +1159,9 @@ int Request::get_status(MPI_Request req, int* flag, MPI_Status * status){ return MPI_SUCCESS; } -int Request::grequest_start( MPI_Grequest_query_function *query_fn, MPI_Grequest_free_function *free_fn, MPI_Grequest_cancel_function *cancel_fn, void *extra_state, MPI_Request *request){ - +int Request::grequest_start(MPI_Grequest_query_function* query_fn, MPI_Grequest_free_function* free_fn, + MPI_Grequest_cancel_function* cancel_fn, void* extra_state, MPI_Request* request) +{ *request = new Request(); (*request)->flags_ |= MPI_REQ_GENERALIZED; (*request)->flags_ |= MPI_REQ_PERSISTENT; @@ -1170,7 +1176,8 @@ int Request::grequest_start( MPI_Grequest_query_function *query_fn, MPI_Grequest return MPI_SUCCESS; } -int Request::grequest_complete( MPI_Request request){ +int Request::grequest_complete(MPI_Request request) +{ if ((!(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex==NULL) return MPI_ERR_REQUEST; request->generalized_funcs->mutex->lock();