X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/2d42b4bde2a305fa165d26f3d81f86e4aa6ff7aa..8bc1da3cc40001eecdeda718441471a29b02523c:/src/smpi/mpi/smpi_request.cpp

diff --git a/src/smpi/mpi/smpi_request.cpp b/src/smpi/mpi/smpi_request.cpp
index c33b284af6..0b0c7408b3 100644
--- a/src/smpi/mpi/smpi_request.cpp
+++ b/src/smpi/mpi/smpi_request.cpp
@@ -36,7 +36,7 @@ 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,
+Request::Request(const void* buf, int count, MPI_Datatype datatype, aid_t src, aid_t dst, int tag, MPI_Comm comm,
                  unsigned flags, MPI_Op op)
     : buf_(const_cast<void*>(buf))
     , old_type_(datatype)
@@ -63,8 +63,6 @@ Request::Request(const void* buf, int count, MPI_Datatype datatype, int src, int
     refcount_ = 1;
   else
     refcount_ = 0;
-  nbc_requests_=nullptr;
-  nbc_requests_size_=0;
   init_buffer(count);
   this->add_f();
 }
@@ -105,7 +103,7 @@ bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request rece
 {
   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",
+  XBT_DEBUG("Trying to match %s of sender src %ld against %ld, 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());
 
@@ -172,64 +170,64 @@ bool Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
 
 void Request::print_request(const char* message) const
 {
-  XBT_VERB("%s  request %p  [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
-       message, this, buf_, size_, src_, dst_, tag_, flags_);
+  XBT_VERB("%s  request %p  [buf = %p, size = %zu, src = %ld, dst = %ld, tag = %d, flags = %x]", 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(),
-                     dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : MPI_PROC_NULL, tag, comm,
+                     dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED | MPI_REQ_BSEND);
 }
 
 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(),
-                     dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : MPI_PROC_NULL, tag, comm,
+                     dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
 }
 
 MPI_Request Request::ssend_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(),
-                     dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : MPI_PROC_NULL, tag, comm,
+                     dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
 }
 
 MPI_Request Request::isend_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(),
-                     dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : MPI_PROC_NULL, tag, comm,
+                     dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
 }
 
-
 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 */
   if(op==MPI_OP_NULL){
-    request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
-                          dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : MPI_PROC_NULL, tag, comm,
+    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,
                           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, comm->group()->actor(src)->get_pid(),
-                          dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : MPI_PROC_NULL, tag, comm,
+    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,
                           MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
-                              MPI_REQ_ACCUMULATE, op);
+                              MPI_REQ_ACCUMULATE,
+                          op);
   }
   return request;
 }
 
 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
 {
-  int source = MPI_PROC_NULL;
+  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)->get_pid();
+    source = comm->group()->actor(src);
   return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
                      source,
                      simgrid::s4u::this_actor::get_pid(), tag, comm,
@@ -240,30 +238,31 @@ MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype,
                                MPI_Op op)
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
-  int source = MPI_PROC_NULL;
+  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)->get_pid();
+    source = comm->group()->actor(src);
   if(op==MPI_OP_NULL){
     request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
-                          dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : MPI_PROC_NULL, tag, comm,
+                          dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, 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, source,
-                          dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : MPI_PROC_NULL, tag, comm,
-                          MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
+    request =
+        new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
+                    dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+                    MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
   }
   return request;
 }
 
 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
 {
-  int source = MPI_PROC_NULL;
+  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)->get_pid();
+    source = comm->group()->actor(src);
   return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
                      source, simgrid::s4u::this_actor::get_pid(), tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
@@ -273,7 +272,7 @@ MPI_Request Request::ibsend(const void *buf, int count, MPI_Datatype datatype, i
 {
   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)->get_pid() : MPI_PROC_NULL, tag, comm,
+                        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();
@@ -284,7 +283,7 @@ MPI_Request Request::isend(const void *buf, int count, MPI_Datatype datatype, in
 {
   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)->get_pid() : MPI_PROC_NULL, tag, comm,
+                        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();
@@ -295,22 +294,21 @@ MPI_Request Request::issend(const void *buf, int count, MPI_Datatype datatype, i
 {
   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)->get_pid() : MPI_PROC_NULL, tag, comm,
+                        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;
 }
 
-
 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 */
-  int source = MPI_PROC_NULL;
+  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)->get_pid();
+    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);
@@ -332,8 +330,8 @@ void Request::bsend(const void *buf, int count, MPI_Datatype datatype, int dst,
 {
   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)->get_pid() : MPI_PROC_NULL, 
-                        tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
+                        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();
@@ -345,8 +343,8 @@ void Request::send(const void *buf, int count, MPI_Datatype datatype, int dst, i
 {
   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)->get_pid() : MPI_PROC_NULL, 
-                        tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
+                        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);
@@ -357,7 +355,7 @@ void Request::ssend(const void *buf, int count, MPI_Datatype datatype, int dst,
 {
   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)->get_pid() : MPI_PROC_NULL, tag, comm,
+                        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)
@@ -370,16 +368,16 @@ void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype
                        void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
                        MPI_Comm comm, MPI_Status * status)
 {
-  int source = MPI_PROC_NULL;
+  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)->get_pid();
-  int destination = dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : MPI_PROC_NULL;
+    source = comm->group()->actor(src);
+  aid_t destination = dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL;
 
   std::array<MPI_Request, 2> requests;
   std::array<MPI_Status, 2> stats;
-  int myid = simgrid::s4u::this_actor::get_pid();
+  aid_t myid = simgrid::s4u::this_actor::get_pid();
   if ((destination == myid) && (source == myid)) {
     Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
     if (status != MPI_STATUS_IGNORE) {
@@ -470,7 +468,7 @@ void Request::start()
       mut->unlock();
   } else { /* the RECV flag was not set, so this is a send */
     const simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
-    xbt_assert(process, "Actor pid=%d is gone??", dst_);
+    xbt_assert(process, "Actor pid=%ld is gone??", dst_);
     if (TRACE_smpi_view_internals())
       TRACE_smpi_send(src_, src_, dst_, tag_, size_);
     this->print_request("New send");
@@ -598,24 +596,17 @@ int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
 
   static int nsleeps = 1;
   int ret = MPI_SUCCESS;
-  
-  // Are we testing a request meant for non blocking collectives ?
-  // If so, test all the subrequests.
-  if ((*request)->nbc_requests_size_>0){
-    ret = testall((*request)->nbc_requests_size_, (*request)->nbc_requests_, flag, MPI_STATUSES_IGNORE);
-    if(*flag){
-      delete[] (*request)->nbc_requests_;
-      (*request)->nbc_requests_size_=0;
-      unref(request);
-    }
-    return ret;
-  }
-  
+
   if(smpi_test_sleep > 0)
     simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
 
   Status::empty(status);
   *flag = 1;
+
+  if ((*request)->flags_ & MPI_REQ_NBC){
+    *flag = finish_nbc_requests(request, 1);
+  }
+
   if (((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) == 0) {
     if ((*request)->action_ != nullptr && ((*request)->flags_ & MPI_REQ_CANCELLED) == 0){
       try{
@@ -732,7 +723,11 @@ int Request::testany(int count, MPI_Request requests[], int *index, int* flag, M
         ret=(requests[*index]->generalized_funcs)->query_fn((requests[*index]->generalized_funcs)->extra_state, mystatus);
       }
 
-        if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT)) 
+      if (requests[*index] != MPI_REQUEST_NULL && requests[*index]->flags_ & MPI_REQ_NBC){
+        *flag = finish_nbc_requests(&requests[*index] , 1);
+      }
+
+      if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT))
           requests[*index] = MPI_REQUEST_NULL;
         XBT_DEBUG("Testany - returning with index %d", *index);
         *flag=1;
@@ -764,7 +759,6 @@ int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status
       int ret = test(&requests[i], pstat, &flag);
       if (flag){
         flag=0;
-        requests[i]=MPI_REQUEST_NULL;
       }else{
         *outflag=0;
       }
@@ -801,9 +795,9 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   static int nsleeps = 1;
   double speed        = s4u::this_actor::get_host()->get_speed();
   double maxrate      = smpi_cfg_iprobe_cpu_usage();
-  auto 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 | MPI_REQ_PROBE);
+  auto request =
+      new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source),
+                  simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PROBE);
   if (smpi_iprobe_sleep > 0) {
     /** Compute the number of flops we will sleep **/
     s4u::this_actor::exec_init(/*nsleeps: See comment above */ nsleeps *
@@ -856,6 +850,39 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   xbt_assert(request == MPI_REQUEST_NULL);
 }
 
+int Request::finish_nbc_requests(MPI_Request* request, int test){
+  int flag = 1;
+  int ret = 0;
+  if(test == 0)
+    ret = waitall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), MPI_STATUSES_IGNORE);
+  else{
+    ret = testall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), &flag, MPI_STATUSES_IGNORE);
+  }
+  if(ret!=MPI_SUCCESS)
+    xbt_die("Failure when waiting on non blocking collective sub-requests");
+  if(flag == 1){
+    XBT_DEBUG("Finishing non blocking collective request with %zu sub-requests", (*request)->nbc_requests_.size());
+    for(auto& req: (*request)->nbc_requests_){
+      if((*request)->buf_!=nullptr && req!=MPI_REQUEST_NULL){//reduce case
+        void * buf=req->buf_;
+        if((*request)->old_type_->flags() & DT_FLAG_DERIVED)
+          buf=req->old_buf_;
+        if(req->flags_ & MPI_REQ_RECV ){
+          if((*request)->op_!=MPI_OP_NULL){
+            int count=(*request)->size_/ (*request)->old_type_->size();
+            (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->old_type_);
+          }
+          smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
+        }
+      }
+      if(req!=MPI_REQUEST_NULL)
+        Request::unref(&req);
+    }
+    (*request)->nbc_requests_.clear();
+  }
+  return flag;
+}
+
 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
 {
   MPI_Request req = *request;
@@ -875,7 +902,7 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
         Status::empty(status);
         status->MPI_SOURCE = MPI_PROC_NULL;
       } else {
-        int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
+        aid_t src          = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
         status->MPI_SOURCE = req->comm_->group()->rank(src);
         status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
         status->MPI_ERROR  = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
@@ -917,8 +944,8 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
   }
 
   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_);
+    aid_t rank       = simgrid::s4u::this_actor::get_pid();
+    aid_t 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){
@@ -969,32 +996,6 @@ int Request::wait(MPI_Request * request, MPI_Status * status)
     (*request)=MPI_REQUEST_NULL;
     return ret;
   }
-  // Are we waiting on a request meant for non blocking collectives ?
-  // If so, wait for all the subrequests.
-  if ((*request)->nbc_requests_size_>0){
-    ret = waitall((*request)->nbc_requests_size_, (*request)->nbc_requests_, MPI_STATUSES_IGNORE);
-    for (int i = 0; i < (*request)->nbc_requests_size_; i++) {
-      if((*request)->buf_!=nullptr && (*request)->nbc_requests_[i]!=MPI_REQUEST_NULL){//reduce case
-        void * buf=(*request)->nbc_requests_[i]->buf_;
-        if((*request)->old_type_->flags() & DT_FLAG_DERIVED)
-          buf=(*request)->nbc_requests_[i]->old_buf_;
-        if((*request)->nbc_requests_[i]->flags_ & MPI_REQ_RECV ){
-          if((*request)->op_!=MPI_OP_NULL){
-            int count=(*request)->size_/ (*request)->old_type_->size();
-            (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->old_type_);
-          }
-          smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
-        }
-      }
-      if((*request)->nbc_requests_[i]!=MPI_REQUEST_NULL)
-        Request::unref(&((*request)->nbc_requests_[i]));
-    }
-    delete[] (*request)->nbc_requests_;
-    (*request)->nbc_requests_size_=0;
-    unref(request);
-    (*request)=MPI_REQUEST_NULL;
-    return ret;
-  }
 
   (*request)->print_request("Waiting");
   if ((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) {
@@ -1031,6 +1032,9 @@ int Request::wait(MPI_Request * request, MPI_Status * status)
   if ((*request)->truncated_)
     ret = MPI_ERR_TRUNCATE;
 
+  if ((*request)->flags_ & MPI_REQ_NBC)
+    finish_nbc_requests(request, 0);
+
   finish_wait(request, status); // may invalidate *request
   if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
     *request = MPI_REQUEST_NULL;
@@ -1057,6 +1061,8 @@ 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)
+            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))
             requests[i] = MPI_REQUEST_NULL; // set to null
@@ -1088,6 +1094,7 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
     }
   }
 
+
   if (index==MPI_UNDEFINED)
     Status::empty(status);
 
@@ -1216,7 +1223,7 @@ int Request::get_status(const Request* req, int* flag, MPI_Status* status)
   *flag=1;
   if(req != MPI_REQUEST_NULL &&
      status != MPI_STATUS_IGNORE) {
-    int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
+    aid_t src          = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
     status->MPI_SOURCE = req->comm_->group()->rank(src);
     status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
     status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
@@ -1253,22 +1260,14 @@ int Request::grequest_complete(MPI_Request request)
   return MPI_SUCCESS;
 }
 
-void Request::set_nbc_requests(MPI_Request* reqs, int size){
-  nbc_requests_size_ = size;
-  if (size > 0) {
+void Request::start_nbc_requests(std::vector<MPI_Request> reqs){
+  if (reqs.size() > 0) {
     nbc_requests_ = reqs;
-  } else {
-    delete[] reqs;
-    nbc_requests_ = nullptr;
+    Request::startall(reqs.size(), reqs.data());
   }
 }
 
-int Request::get_nbc_requests_size() const
-{
-  return nbc_requests_size_;
-}
-
-MPI_Request* Request::get_nbc_requests() const
+std::vector<MPI_Request> Request::get_nbc_requests() const
 {
   return nbc_requests_;
 }