X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/9322977e4345ed358e0d3b2e2705926038b84042..2d37e348a09783cda723c7019640ee69de168324:/src/smpi/mpi/smpi_request.cpp diff --git a/src/smpi/mpi/smpi_request.cpp b/src/smpi/mpi/smpi_request.cpp index 8e9f37fd6c..dc751be4bf 100644 --- a/src/smpi/mpi/smpi_request.cpp +++ b/src/smpi/mpi/smpi_request.cpp @@ -1,24 +1,27 @@ -/* 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_request.hpp" -#include "SmpiHost.hpp" #include "mc/mc.h" #include "private.hpp" +#include "simgrid/Exception.hpp" +#include "simgrid/s4u/Exec.hpp" #include "smpi_comm.hpp" #include "smpi_datatype.hpp" +#include "smpi_host.hpp" #include "smpi_op.hpp" -#include "smpi_process.hpp" #include "src/kernel/activity/CommImpl.hpp" #include "src/mc/mc_replay.hpp" #include "src/simix/ActorImpl.hpp" +#include "src/smpi/include/smpi_actor.hpp" +#include "xbt/config.hpp" #include -XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (reques)"); +XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (request)"); static simgrid::config::Flag smpi_iprobe_sleep( "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4); @@ -37,14 +40,14 @@ Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, { void *old_buf = nullptr; // FIXME Handle the case of a partial shared malloc. - if ((((flags & RECV) != 0) && ((flags & ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) { + 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 = buf; if (count==0){ buf_ = nullptr; }else { buf_ = xbt_malloc(count*datatype->size()); - if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) { + if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & MPI_REQ_SEND) != 0)) { datatype->serialize(old_buf, buf_, count); } } @@ -61,11 +64,12 @@ Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, truncated_ = 0; real_size_ = 0; real_tag_ = 0; - if (flags & PERSISTENT) + if (flags & MPI_REQ_PERSISTENT) refcount_ = 1; else refcount_ = 0; op_ = MPI_REPLACE; + cancelled_ = 0; } MPI_Comm Request::comm(){ @@ -92,6 +96,10 @@ int Request::detached(){ return detached_; } +MPI_Datatype Request::type(){ + return old_type_; +} + size_t Request::size(){ return size_; } @@ -100,11 +108,18 @@ size_t Request::real_size(){ return real_size_; } +void Request::ref(){ + refcount_++; +} + void Request::unref(MPI_Request* request) { if((*request) != MPI_REQUEST_NULL){ (*request)->refcount_--; - if((*request)->refcount_<0) xbt_die("wrong refcount"); + if((*request)->refcount_ < 0) { + (*request)->print_request("wrong refcount"); + xbt_die("Whoops, wrong refcount"); + } if((*request)->refcount_==0){ Datatype::unref((*request)->old_type_); Comm::unref((*request)->comm_); @@ -119,7 +134,7 @@ void Request::unref(MPI_Request* request) } } -int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored) +int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*) { MPI_Request ref = static_cast(a); MPI_Request req = static_cast(b); @@ -138,12 +153,14 @@ int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* i ref->truncated_ = 1; if(req->detached_==1) 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) + req->cancelled_=-1;//mark as uncancellable XBT_DEBUG("match succeeded"); return 1; }else return 0; } -int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored) +int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*) { MPI_Request ref = static_cast(a); MPI_Request req = static_cast(b); @@ -161,6 +178,8 @@ int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl* i req->truncated_ = 1; if(ref->detached_==1) 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 uncancellable XBT_DEBUG("match succeeded"); return 1; } else @@ -178,20 +197,23 @@ void Request::print_request(const char *message) MPI_Request Request::send_init(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::Actor::self()->getPid() - 1, - comm->group()->actor(dst)->getPid() - 1, tag, comm, PERSISTENT | SEND | PREPARED); + 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_Request Request::ssend_init(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::Actor::self()->getPid() - 1, - comm->group()->actor(dst)->getPid() - 1, tag, comm, PERSISTENT | SSEND | SEND | PREPARED); + 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_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED); } MPI_Request Request::isend_init(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::Actor::self()->getPid() - 1, - comm->group()->actor(dst)->getPid() - 1, tag, comm, PERSISTENT | ISEND | SEND | PREPARED); + 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_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED); } @@ -200,11 +222,14 @@ MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, { MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */ if(op==MPI_OP_NULL){ - request = new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag, - comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED); + request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(), + comm->group()->actor(dst)->get_pid(), tag, comm, + MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED); }else{ - request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag, - comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE); + request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(), + comm->group()->actor(dst)->get_pid(), tag, comm, + MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED | + MPI_REQ_ACCUMULATE); request->op_ = op; } return request; @@ -213,8 +238,9 @@ MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) { return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, - src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->getPid() - 1, - simgrid::s4u::Actor::self()->getPid() - 1, tag, comm, PERSISTENT | RECV | PREPARED); + src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(), + simgrid::s4u::this_actor::get_pid(), tag, comm, + MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED); } MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, @@ -222,11 +248,13 @@ MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, { MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */ if(op==MPI_OP_NULL){ - request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag, comm, - RMA | NON_PERSISTENT | RECV | PREPARED); + request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(), + comm->group()->actor(dst)->get_pid(), tag, comm, + MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED); }else{ - request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag, comm, - RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE); + request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(), + comm->group()->actor(dst)->get_pid(), tag, comm, + MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE); request->op_ = op; } return request; @@ -235,15 +263,17 @@ MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) { return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, - src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->getPid() - 1, - simgrid::s4u::Actor::self()->getPid() - 1, tag, comm, PERSISTENT | RECV | PREPARED); + src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(), + simgrid::s4u::this_actor::get_pid(), tag, comm, + MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED); } MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */ - request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::Actor::self()->getPid() - 1, - comm->group()->actor(dst)->getPid() - 1, tag, comm, NON_PERSISTENT | ISEND | SEND); + request = 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_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND); request->start(); return request; } @@ -251,8 +281,9 @@ MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */ - request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::Actor::self()->getPid() - 1, - comm->group()->actor(dst)->getPid() - 1, tag, comm, NON_PERSISTENT | ISEND | SSEND | SEND); + request = 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_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND); request->start(); return request; } @@ -262,8 +293,8 @@ MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, { MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */ request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, - src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->getPid() - 1, - simgrid::s4u::Actor::self()->getPid() - 1, tag, comm, NON_PERSISTENT | RECV); + src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(), + simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV); request->start(); return request; } @@ -279,8 +310,8 @@ void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */ - request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::Actor::self()->getPid() - 1, - comm->group()->actor(dst)->getPid() - 1, tag, comm, NON_PERSISTENT | SEND); + request = 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_NON_PERSISTENT | MPI_REQ_SEND); request->start(); wait(&request, MPI_STATUS_IGNORE); @@ -290,8 +321,9 @@ void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */ - request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::Actor::self()->getPid() - 1, - comm->group()->actor(dst)->getPid() - 1, tag, comm, NON_PERSISTENT | SSEND | SEND); + request = 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_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND); request->start(); wait(&request,MPI_STATUS_IGNORE); @@ -304,10 +336,16 @@ void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int d { MPI_Request requests[2]; MPI_Status stats[2]; - unsigned int myid = simgrid::s4u::Actor::self()->getPid() - 1; - if ((comm->group()->actor(dst)->getPid()-1 == myid) && (comm->group()->actor(src)->getPid()-1 == myid)){ - Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype); - return; + int myid = simgrid::s4u::this_actor::get_pid(); + if ((comm->group()->actor(dst)->get_pid() == myid) && (comm->group()->actor(src)->get_pid() == myid)) { + Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype); + if (status != MPI_STATUS_IGNORE) { + status->MPI_SOURCE = src; + status->MPI_TAG = recvtag; + status->MPI_ERROR = MPI_SUCCESS; + status->count = sendcount * sendtype->size(); + } + return; } requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm); requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm); @@ -326,25 +364,24 @@ void Request::start() smx_mailbox_t mailbox; xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication"); - flags_ &= ~PREPARED; - flags_ &= ~FINISHED; - refcount_++; + flags_ &= ~MPI_REQ_PREPARED; + flags_ &= ~MPI_REQ_FINISHED; + this->ref(); - if ((flags_ & RECV) != 0) { + if ((flags_ & MPI_REQ_RECV) != 0) { this->print_request("New recv"); - simgrid::smpi::Process* process = smpi_process_remote(simgrid::s4u::Actor::byPid(dst_+1)); + simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_)); - int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh"); + int async_small_thresh = simgrid::config::get_value("smpi/async-small-thresh"); xbt_mutex_t mut = process->mailboxes_mutex(); - if (async_small_thresh != 0 || (flags_ & RMA) != 0) + if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0) xbt_mutex_acquire(mut); - if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) { + if (async_small_thresh == 0 && (flags_ & MPI_REQ_RMA) == 0) { mailbox = process->mailbox(); - } - else if (((flags_ & RMA) != 0) || static_cast(size_) < async_small_thresh) { + } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast(size_) < 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(); @@ -378,14 +415,14 @@ 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_; action_ = simcall_comm_irecv( - process->process()->getImpl(), mailbox, buf_, &real_size_, &match_recv, + process->get_actor()->get_impl(), mailbox, buf_, &real_size_, &match_recv, 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_ & RMA) != 0 ) + if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0) xbt_mutex_release(mut); } else { /* the RECV flag was not set, so this is a send */ - simgrid::smpi::Process* process = smpi_process_remote(simgrid::s4u::Actor::byPid(dst_+1)); + simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_)); int rank = src_; if (TRACE_smpi_view_internals()) { TRACE_smpi_send(rank, rank, dst_, tag_, size_); @@ -393,19 +430,21 @@ void Request::start() this->print_request("New send"); void* buf = buf_; - if ((flags_ & SSEND) == 0 && ( (flags_ & RMA) != 0 - || static_cast(size_) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) { + if ((flags_ & MPI_REQ_SSEND) == 0 && + ((flags_ & MPI_REQ_RMA) != 0 || + static_cast(size_) < simgrid::config::get_value("smpi/send-is-detached-thresh"))) { void *oldbuf = nullptr; detached_ = 1; XBT_DEBUG("Send request %p is detached", this); - refcount_++; + this->ref(); if (not(old_type_->flags() & DT_FLAG_DERIVED)) { oldbuf = buf_; if (not process->replaying() && oldbuf != nullptr && size_ != 0) { - if ((smpi_privatize_global_variables != 0) && (static_cast(buf_) >= smpi_data_exe_start) && + if ((smpi_privatize_global_variables != 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 "); - smpi_switch_data_segment(src_); + smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_)); } buf = xbt_malloc(size_); memcpy(buf,oldbuf,size_); @@ -416,11 +455,11 @@ void Request::start() //if we are giving back the control to the user without waiting for completion, we have to inject timings double sleeptime = 0.0; - if (detached_ != 0 || ((flags_ & (ISEND | SSEND)) != 0)) { // issend should be treated as isend + if (detached_ != 0 || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend // isend and send timings may be different - sleeptime = ((flags_ & ISEND) != 0) - ? simgrid::s4u::Actor::self()->getHost()->extension()->oisend(size_) - : simgrid::s4u::Actor::self()->getHost()->extension()->osend(size_); + sleeptime = ((flags_ & MPI_REQ_ISEND) != 0) + ? simgrid::s4u::Actor::self()->get_host()->extension()->oisend(size_) + : simgrid::s4u::Actor::self()->get_host()->extension()->osend(size_); } if(sleeptime > 0.0){ @@ -428,21 +467,21 @@ void Request::start() XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime); } - int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh"); + int async_small_thresh = simgrid::config::get_value("smpi/async-small-thresh"); xbt_mutex_t mut=process->mailboxes_mutex(); - if (async_small_thresh != 0 || (flags_ & RMA) != 0) + if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0) xbt_mutex_acquire(mut); - if (not(async_small_thresh != 0 || (flags_ & RMA) != 0)) { + if (not(async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)) { mailbox = process->mailbox(); - } else if (((flags_ & RMA) != 0) || static_cast(size_) < async_small_thresh) { // eager mode + } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast(size_) < async_small_thresh) { // eager mode mailbox = process->mailbox(); XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox); smx_activity_t action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast(this)); if (action == nullptr) { - if ((flags_ & SSEND) == 0){ + if ((flags_ & MPI_REQ_SSEND) == 0) { mailbox = process->mailbox_small(); XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox); } else { @@ -465,7 +504,7 @@ 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_; action_ = simcall_comm_isend( - simgrid::s4u::Actor::byPid(src_ + 1)->getImpl(), mailbox, size_, -1.0, buf, real_size_, &match_send, + simgrid::s4u::Actor::by_pid(src_)->get_impl(), mailbox, size_, -1.0, buf, real_size_, &match_send, &xbt_free_f, // how to free the userdata if a detached send fails not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this, // detach if msg size < eager/rdv switch limit @@ -475,7 +514,7 @@ void Request::start() /* FIXME: detached sends are not traceable (action_ == nullptr) */ if (action_ != nullptr) simcall_set_category(action_, TRACE_internal_smpi_get_category()); - if (async_small_thresh != 0 || ((flags_ & RMA)!=0)) + if (async_small_thresh != 0 || ((flags_ & MPI_REQ_RMA) != 0)) xbt_mutex_release(mut); } } @@ -490,6 +529,14 @@ void Request::startall(int count, MPI_Request * requests) } } +void Request::cancel() +{ + if(cancelled_!=-1) + cancelled_=1; + if (this->action_ != nullptr) + (boost::static_pointer_cast(this->action_))->cancel(); +} + int Request::test(MPI_Request * request, MPI_Status * status) { //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before) // to avoid deadlocks if used as a break condition, such as @@ -502,15 +549,20 @@ int Request::test(MPI_Request * request, MPI_Status * status) { Status::empty(status); int flag = 1; - if (((*request)->flags_ & PREPARED) == 0) { - if ((*request)->action_ != nullptr) - flag = simcall_comm_test((*request)->action_); + if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) { + if ((*request)->action_ != nullptr){ + try{ + flag = simcall_comm_test((*request)->action_); + }catch (xbt_ex& e) { + return 0; + } + } if (flag) { finish_wait(request,status); nsleeps=1;//reset the number of sleeps we will do next time - if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & PERSISTENT) == 0) + if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0) *request = MPI_REQUEST_NULL; - } else if (xbt_cfg_get_boolean("smpi/grow-injected-times")){ + } else if (simgrid::config::get_value("smpi/grow-injected-times")) { nsleeps++; } } @@ -531,7 +583,7 @@ int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Sta count++; if (status != MPI_STATUSES_IGNORE) status[i] = *pstat; - if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & NON_PERSISTENT)) + if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT)) requests[i] = MPI_REQUEST_NULL; } } else { @@ -555,7 +607,7 @@ int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * std::vector map; /** Maps all matching comms back to their location in requests **/ for(i = 0; i < count; i++) { - if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & PREPARED)) { + if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) { comms.push_back(requests[i]->action_); map.push_back(i); } @@ -565,14 +617,18 @@ int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * static int nsleeps = 1; if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep); - - i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches! + try{ + i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches! + }catch (xbt_ex& e) { + return 0; + } + if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches) *index = map[i]; finish_wait(&requests[*index],status); flag = 1; nsleeps = 1; - if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & NON_PERSISTENT)) { + if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT)) { requests[*index] = MPI_REQUEST_NULL; } } else { @@ -593,7 +649,7 @@ int Request::testall(int count, MPI_Request requests[], MPI_Status status[]) MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat; int flag=1; for(int i=0; iflags_ & PREPARED)) { + if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) { if (test(&requests[i], pstat)!=1){ flag=0; }else{ @@ -625,26 +681,28 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it // This can speed up the execution of certain applications by an order of magnitude, such as HPL static int nsleeps = 1; - double speed = simgrid::s4u::Actor::self()->getHost()->getSpeed(); - double maxrate = xbt_cfg_get_double("smpi/iprobe-cpu-usage"); - MPI_Request request = new Request( - nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source)->getPid() - 1, - simgrid::s4u::Actor::self()->getPid() - 1, tag, comm, PERSISTENT | RECV); + double speed = s4u::this_actor::get_host()->get_speed(); + double maxrate = simgrid::config::get_value("smpi/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); if (smpi_iprobe_sleep > 0) { - smx_activity_t iprobe_sleep = simcall_execution_start( - "iprobe", /* flops to executek*/ nsleeps * smpi_iprobe_sleep * speed * maxrate, /* priority */ 1.0, - /* performance bound */ maxrate * speed, smpi_process()->process()->getImpl()->host); - simcall_execution_wait(iprobe_sleep); + /** Compute the number of flops we will sleep **/ + s4u::this_actor::exec_init(/*nsleeps: See comment above */ nsleeps * + /*(in seconds)*/ smpi_iprobe_sleep * speed * maxrate) + ->set_name("iprobe") + ->start() + ->wait(); } // behave like a receive, but don't do it smx_mailbox_t mailbox; request->print_request("New iprobe"); // We have to test both mailboxes as we don't know if we will receive one one or another - if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){ - mailbox = smpi_process()->mailbox_small(); - XBT_DEBUG("Trying to probe the perm recv mailbox"); - request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast(request)); + if (simgrid::config::get_value("smpi/async-small-thresh") > 0) { + mailbox = smpi_process()->mailbox_small(); + XBT_DEBUG("Trying to probe the perm recv mailbox"); + request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast(request)); } if (request->action_ == nullptr){ @@ -654,11 +712,10 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* } if (request->action_ != nullptr){ - simgrid::kernel::activity::CommImplPtr sync_comm = - boost::static_pointer_cast(request->action_); - MPI_Request req = static_cast(sync_comm->src_data); + kernel::activity::CommImplPtr sync_comm = boost::static_pointer_cast(request->action_); + MPI_Request req = static_cast(sync_comm->src_data_); *flag = 1; - if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) { + if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) { status->MPI_SOURCE = comm->group()->rank(req->src_); status->MPI_TAG = req->tag_; status->MPI_ERROR = MPI_SUCCESS; @@ -668,7 +725,7 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* } else { *flag = 0; - if (xbt_cfg_get_boolean("smpi/grow-injected-times")) + if (simgrid::config::get_value("smpi/grow-injected-times")) nsleeps++; } unref(&request); @@ -678,8 +735,17 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status) { MPI_Request req = *request; Status::empty(status); + + if (req->cancelled_==1){ + if (status!=MPI_STATUS_IGNORE) + status->cancelled=1; + if(req->detached_sender_ != nullptr) + unref(&(req->detached_sender_)); + unref(request); + return; + } - if (not((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)) { + if (not((req->detached_ != 0) && ((req->flags_ & MPI_REQ_SEND) != 0)) && ((req->flags_ & MPI_REQ_PREPARED) == 0)) { if(status != MPI_STATUS_IGNORE) { int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_; status->MPI_SOURCE = req->comm_->group()->rank(src); @@ -693,66 +759,72 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status) MPI_Datatype datatype = req->old_type_; // FIXME Handle the case of a partial shared malloc. - if (((req->flags_ & ACCUMULATE) != 0) || + 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 != 0 && + if (not smpi_process()->replaying() && smpi_privatize_global_variables != 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 "); - smpi_switch_data_segment(simgrid::s4u::Actor::self()->getPid() - 1); + smpi_switch_data_segment(simgrid::s4u::Actor::self()); } if(datatype->flags() & DT_FLAG_DERIVED){ // This part handles the problem of non-contignous memory the unserialization at the reception - if((req->flags_ & RECV) && datatype->size()!=0) + 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_); - }else if(req->flags_ & 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_); + } 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_); } } } - if (TRACE_smpi_view_internals() && ((req->flags_ & RECV) != 0)){ - int rank = simgrid::s4u::Actor::self()->getPid() - 1; + if (TRACE_smpi_view_internals() && ((req->flags_ & MPI_REQ_RECV) != 0)) { + int rank = simgrid::s4u::this_actor::get_pid(); int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_); TRACE_smpi_recv(src_traced, rank,req->tag_); } if(req->detached_sender_ != nullptr){ //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0 double sleeptime = - simgrid::s4u::Actor::self()->getHost()->extension()->orecv(req->real_size()); + simgrid::s4u::Actor::self()->get_host()->extension()->orecv(req->real_size()); if(sleeptime > 0.0){ simcall_process_sleep(sleeptime); XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime); } unref(&(req->detached_sender_)); } - if(req->flags_ & PERSISTENT) + if (req->flags_ & MPI_REQ_PERSISTENT) req->action_ = nullptr; - req->flags_ |= FINISHED; + req->flags_ |= MPI_REQ_FINISHED; unref(request); } void Request::wait(MPI_Request * request, MPI_Status * status) { (*request)->print_request("Waiting"); - if ((*request)->flags_ & PREPARED) { + if ((*request)->flags_ & MPI_REQ_PREPARED) { Status::empty(status); return; } - if ((*request)->action_ != nullptr) - // this is not a detached send - simcall_comm_wait((*request)->action_, -1.0); + if ((*request)->action_ != nullptr){ + try{ + // this is not a detached send + simcall_comm_wait((*request)->action_, -1.0); + }catch (xbt_ex& e) { + XBT_VERB("Request cancelled"); + } + } + finish_wait(request,status); - if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & NON_PERSISTENT)!=0)) + if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0)) *request = MPI_REQUEST_NULL; } @@ -770,8 +842,8 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status) int *map = xbt_new(int, count); XBT_DEBUG("Wait for one of %d", count); for(int i = 0; i < count; i++) { - if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED) && - not(requests[i]->flags_ & FINISHED)) { + if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED) && + not(requests[i]->flags_ & MPI_REQ_FINISHED)) { if (requests[i]->action_ != nullptr) { XBT_DEBUG("Waiting any %p ", requests[i]); intrusive_ptr_add_ref(requests[i]->action_.get()); @@ -783,7 +855,7 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status) size = 0; // so we free the dynar but don't do the waitany call index = i; finish_wait(&requests[i], status); // cleanup if refcount = 0 - if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT)) + if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT)) requests[i] = MPI_REQUEST_NULL; // set to null break; } @@ -791,16 +863,23 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status) } if (size > 0) { XBT_DEBUG("Enter waitany for %lu comms", xbt_dynar_length(&comms)); - int i = simcall_comm_waitany(&comms, -1); + int i=MPI_UNDEFINED; + try{ + // this is not a detached send + i = simcall_comm_waitany(&comms, -1); + }catch (xbt_ex& e) { + XBT_INFO("request %d cancelled ",i); + return i; + } // not MPI_UNDEFINED, as this is a simix return code if (i != -1) { index = map[i]; //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly. if ((requests[index] == MPI_REQUEST_NULL) || - (not((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))) { + (not((requests[index]->flags_ & MPI_REQ_ACCUMULATE) && (requests[index]->flags_ & MPI_REQ_RECV)))) { finish_wait(&requests[index],status); - if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT)) + if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT)) requests[index] = MPI_REQUEST_NULL; } } @@ -831,7 +910,8 @@ int Request::waitall(int count, MPI_Request requests[], MPI_Status status[]) //tag invalid requests in the set if (status != MPI_STATUSES_IGNORE) { for (int c = 0; c < count; c++) { - if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || (requests[c]->flags_ & PREPARED)) { + if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || + (requests[c]->flags_ & MPI_REQ_PREPARED)) { Status::empty(&status[c]); } else if (requests[c]->src_ == MPI_PROC_NULL) { Status::empty(&status[c]); @@ -845,14 +925,14 @@ int Request::waitall(int count, MPI_Request requests[], MPI_Status status[]) index = c; } else { index = waitany(count, (MPI_Request*)requests, pstat); + if (index == MPI_UNDEFINED) break; - if (requests[index] != MPI_REQUEST_NULL - && (requests[index]->flags_ & RECV) - && (requests[index]->flags_ & ACCUMULATE)) + if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_RECV) && + (requests[index]->flags_ & MPI_REQ_ACCUMULATE)) accumulates.push_back(requests[index]); - if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT)) + if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT)) requests[index] = MPI_REQUEST_NULL; } if (status != MPI_STATUSES_IGNORE) { @@ -886,8 +966,8 @@ int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Sta if(status != MPI_STATUSES_IGNORE) { status[index] = *pstat; } - if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT)) - requests[index] = MPI_REQUEST_NULL; + if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT)) + requests[index] = MPI_REQUEST_NULL; }else{ return MPI_UNDEFINED; }