X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/0199ba108d66c94df94e4f044994e79efdece4b1..ea74f5d95928a521a588737e81f1de94eef25d19:/src/smpi/mpi/smpi_request.cpp

diff --git a/src/smpi/mpi/smpi_request.cpp b/src/smpi/mpi/smpi_request.cpp
index 72c4f99e74..2b54eacf68 100644
--- a/src/smpi/mpi/smpi_request.cpp
+++ b/src/smpi/mpi/smpi_request.cpp
@@ -1,4 +1,4 @@
-/* Copyright (c) 2007-2021. The SimGrid Team. All rights reserved.          */
+/* Copyright (c) 2007-2022. 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. */
@@ -57,6 +57,10 @@ Request::Request(const void* buf, int count, MPI_Datatype datatype, aid_t src, a
   detached_        = false;
   detached_sender_ = nullptr;
   real_src_        = 0;
+  // get src_host if it's available (src is valid)
+  auto src_process = simgrid::s4u::Actor::by_pid(src);
+  if (src_process)
+    src_host_ = src_process->get_host();
   truncated_       = false;
   unmatched_types_ = false;
   real_size_       = 0;
@@ -132,8 +136,10 @@ bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request rece
        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)
+    if (receiver->src_ == MPI_ANY_SOURCE) {
       receiver->real_src_ = sender->src_;
+      receiver->src_host_ = sender->src_host_;
+    }
     if (receiver->tag_ == MPI_ANY_TAG)
       receiver->real_tag_ = sender->tag_;
     if ((receiver->flags_ & MPI_REQ_PROBE) == 0 ){
@@ -161,7 +167,8 @@ bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request rece
 void Request::init_buffer(int count){
 // 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)) || (type_->flags() & DT_FLAG_DERIVED)) {
+  if (not smpi_process()->replaying() &&
+     ((((flags_ & MPI_REQ_RECV) != 0) && ((flags_ & MPI_REQ_ACCUMULATE) != 0)) || (type_->flags() & DT_FLAG_DERIVED))) {
     // This part handles the problem of non-contiguous memory
     old_buf_ = buf_;
     if (count==0){
@@ -227,7 +234,7 @@ MPI_Request Request::isend_init(const void *buf, int count, MPI_Datatype datatyp
 MPI_Request Request::rma_send_init(const void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
                                MPI_Op op)
 {
-  MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
+  MPI_Request request;
   if(op==MPI_OP_NULL){
     request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
                           dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
@@ -258,12 +265,12 @@ MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int
 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
                                MPI_Op op)
 {
-  MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   aid_t source        = MPI_PROC_NULL;
   if (src == MPI_ANY_SOURCE)
     source = MPI_ANY_SOURCE;
   else if (src != MPI_PROC_NULL)
     source = comm->group()->actor(src);
+  MPI_Request request;
   if(op==MPI_OP_NULL){
     request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
                           dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
@@ -291,10 +298,9 @@ MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int
 
 MPI_Request Request::ibsend(const 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::this_actor::get_pid(),
-                        dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
-                        MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
+  auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                             dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+                             MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
   if(dst != MPI_PROC_NULL)
     request->start();
   return request;
@@ -302,10 +308,9 @@ MPI_Request Request::ibsend(const void *buf, int count, MPI_Datatype datatype, i
 
 MPI_Request Request::isend(const 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::this_actor::get_pid(),
-                        dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
-                        MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
+  auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                             dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+                             MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
   if(dst != MPI_PROC_NULL)
     request->start();
   return request;
@@ -313,10 +318,9 @@ MPI_Request Request::isend(const void *buf, int count, MPI_Datatype datatype, in
 
 MPI_Request Request::issend(const 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::this_actor::get_pid(),
-                        dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
-                        MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
+  auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                             dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+                             MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
   if(dst != MPI_PROC_NULL)
     request->start();
   return request;
@@ -324,15 +328,13 @@ MPI_Request Request::issend(const void *buf, int count, MPI_Datatype datatype, i
 
 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
 {
-  MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   aid_t source        = MPI_PROC_NULL;
   if (src == MPI_ANY_SOURCE)
     source = MPI_ANY_SOURCE;
   else if (src != MPI_PROC_NULL)
     source = comm->group()->actor(src);
-  request             = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
-                        source, simgrid::s4u::this_actor::get_pid(), tag, comm, 
-                        MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
+  auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
+                             simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
   if(src != MPI_PROC_NULL)
     request->start();
   return request;
@@ -340,8 +342,7 @@ MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src,
 
 int Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
 {
-  MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
-  request = irecv(buf, count, datatype, src, tag, comm);
+  MPI_Request request = irecv(buf, count, datatype, src, tag, comm);
   int retval = wait(&request,status);
   request = nullptr;
   return retval;
@@ -349,10 +350,9 @@ int Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag,
 
 void Request::bsend(const 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::this_actor::get_pid(),
-                        dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
-                        MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
+  auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                             dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+                             MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
 
   if(dst != MPI_PROC_NULL)
    request->start();
@@ -362,10 +362,9 @@ void Request::bsend(const void *buf, int count, MPI_Datatype datatype, int dst,
 
 void Request::send(const 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::this_actor::get_pid(),
-                        dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
-                        MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
+  auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                             dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+                             MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
   if(dst != MPI_PROC_NULL)
    request->start();
   wait(&request, MPI_STATUS_IGNORE);
@@ -374,10 +373,9 @@ void Request::send(const void *buf, int count, MPI_Datatype datatype, int dst, i
 
 void Request::ssend(const 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::this_actor::get_pid(),
-                        dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
-                        MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
+  auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                             dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+                             MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
 
   if(dst != MPI_PROC_NULL)
    request->start();
@@ -521,9 +519,13 @@ void Request::start()
     double sleeptime = 0.0;
     if (detached_ || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
       // isend and send timings may be different
-      sleeptime = ((flags_ & MPI_REQ_ISEND) != 0)
-                      ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(size_)
-                      : simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->osend(size_);
+      sleeptime =
+          ((flags_ & MPI_REQ_ISEND) != 0)
+              ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(
+                    size_, simgrid::s4u::Actor::by_pid(src_)->get_host(), simgrid::s4u::Actor::by_pid(dst_)->get_host())
+              : simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->osend(
+                    size_, simgrid::s4u::Actor::by_pid(src_)->get_host(),
+                    simgrid::s4u::Actor::by_pid(dst_)->get_host());
     }
 
     if(sleeptime > 0.0){
@@ -697,13 +699,12 @@ int Request::testany(int count, MPI_Request requests[], int *index, int* flag, M
   std::vector<simgrid::kernel::activity::CommImpl*> comms;
   comms.reserve(count);
 
-  int i;
   *flag = 0;
   int ret = MPI_SUCCESS;
   *index = MPI_UNDEFINED;
 
   std::vector<int> map; /** Maps all matching comms back to their location in requests **/
-  for(i = 0; i < count; i++) {
+  for (int i = 0; i < count; i++) {
     if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
       comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
       map.push_back(i);
@@ -714,13 +715,14 @@ int Request::testany(int count, MPI_Request requests[], int *index, int* flag, M
     static int nsleeps = 1;
     if(smpi_test_sleep > 0)
       simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
+    ssize_t i;
     try{
       i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
     } catch (const Exception&) {
       XBT_DEBUG("Exception in testany");
       return 0;
     }
-    
+
     if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
       *index = map[i];
       if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED) &&
@@ -779,7 +781,7 @@ int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status
       }else{
         *outflag=0;
       }
-      if (ret != MPI_SUCCESS) 
+      if (ret != MPI_SUCCESS)
         error = 1;
     }else{
       Status::empty(pstat);
@@ -788,9 +790,9 @@ int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status
       status[i] = *pstat;
     }
   }
-  if(error==1) 
+  if (error == 1)
     return MPI_ERR_IN_STATUS;
-  else 
+  else
     return MPI_SUCCESS;
 }
 
@@ -821,11 +823,11 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
                                /*(seconds * flop/s -> total flops)*/ smpi_iprobe_sleep * speed * maxrate)
         ->set_name("iprobe")
         /* Not the entire CPU can be used when iprobing: This is important for
-         * the energy consumption caused by polling with iprobes. 
+         * the energy consumption caused by polling with iprobes.
          * Note also that the number of flops that was
          * computed above contains a maxrate factor and is hence reduced (maxrate < 1)
          */
-        ->set_bound(maxrate*speed)
+        ->set_bound(maxrate * speed)
         ->start()
         ->wait();
   }
@@ -933,9 +935,9 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
       MPI_Datatype datatype = req->type_;
 
       // 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_switch_data_segment(simgrid::s4u::Actor::self(), req->old_buf_))
+      if (not smpi_process()->replaying() &&
+        (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED))) {
+        if (smpi_switch_data_segment(simgrid::s4u::Actor::self(), req->old_buf_))
           XBT_VERB("Privatization : We are unserializing to a zone in global memory  Switch data segment ");
 
         if(datatype->flags() & DT_FLAG_DERIVED){
@@ -963,8 +965,9 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
   }
   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()->get_host()->extension<simgrid::smpi::Host>()->orecv(req->real_size());
+    simgrid::s4u::Host* dst_host = simgrid::s4u::Actor::by_pid(req->dst_)->get_host();
+    double sleeptime             = simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(
+        req->real_size(), req->src_host_, dst_host);
     if (sleeptime > 0.0) {
       simgrid::s4u::this_actor::sleep_for(sleeptime);
       XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
@@ -1079,7 +1082,7 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
           // This is a finished detached request, let's return this one
           comms.clear(); // don't do the waitany call afterwards
           index = i;
-          if (requests[index] != MPI_REQUEST_NULL && (requests[index])->flags_ & MPI_REQ_NBC)
+          if (requests[index]->flags_ & MPI_REQ_NBC)
             finish_nbc_requests(&requests[index], 0);
           finish_wait(&requests[i], status); // cleanup if refcount = 0
           if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
@@ -1090,7 +1093,7 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
     }
     if (not comms.empty()) {
       XBT_DEBUG("Enter waitany for %zu comms", comms.size());
-      int i;
+      ssize_t i;
       try{
         i = simcall_comm_waitany(comms.data(), comms.size(), -1);
       } catch (const CancelException&) {
@@ -1188,8 +1191,7 @@ int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Sta
   indices[count] = index;
   count++;
   for (int i = 0; i < incount; i++) {
-    if (i!=index && requests[i] != MPI_REQUEST_NULL 
-        && not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
+    if (i != index && requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
       test(&requests[i], pstat,&flag);
       if (flag==1){
         indices[count] = i;