X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/f23be1900e1e6d0297407b0916cc37c97b1df269..989042c42c15e746278c98212cd16e655d03f182:/src/smpi/mpi/smpi_request.cpp

diff --git a/src/smpi/mpi/smpi_request.cpp b/src/smpi/mpi/smpi_request.cpp
index 609e0366aa..b782759c68 100644
--- a/src/smpi/mpi/smpi_request.cpp
+++ b/src/smpi/mpi/smpi_request.cpp
@@ -1,15 +1,16 @@
-/* Copyright (c) 2007-2019. The SimGrid Team. All rights reserved.          */
+/* Copyright (c) 2007-2021. 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 "simgrid/s4u/Mutex.hpp"
-#include "simgrid/s4u/ConditionVariable.hpp"
+
 #include "smpi_request.hpp"
 
 #include "mc/mc.h"
 #include "private.hpp"
 #include "simgrid/Exception.hpp"
 #include "simgrid/s4u/Exec.hpp"
+#include "simgrid/s4u/Mutex.hpp"
+#include "simgrid/s4u/ConditionVariable.hpp"
 #include "smpi_comm.hpp"
 #include "smpi_datatype.hpp"
 #include "smpi_host.hpp"
@@ -17,10 +18,9 @@
 #include "src/kernel/activity/CommImpl.hpp"
 #include "src/mc/mc_replay.hpp"
 #include "src/smpi/include/smpi_actor.hpp"
-#include "xbt/config.hpp"
-
 
 #include <algorithm>
+#include <array>
 
 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (request)");
 
@@ -36,26 +36,18 @@ extern void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*
 namespace simgrid{
 namespace smpi{
 
-Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags, MPI_Op op)
-    : buf_(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, aid_t src, aid_t dst, int tag, MPI_Comm comm,
+                 unsigned flags, MPI_Op op)
+    : buf_(const_cast<void*>(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 = 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 unserialisation at the reception)
-  old_buf_  = old_buf;
-  size_ = datatype->size() * count;
   datatype->ref();
   comm_->ref();
   if(op != MPI_REPLACE && op != MPI_OP_NULL)
@@ -71,10 +63,10 @@ Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst,
     refcount_ = 1;
   else
     refcount_ = 0;
-  cancelled_ = 0;
-  generalized_funcs=nullptr;
   nbc_requests_=nullptr;
   nbc_requests_size_=0;
+  init_buffer(count);
+  this->add_f();
 }
 
 void Request::ref(){
@@ -83,138 +75,163 @@ void Request::ref(){
 
 void Request::unref(MPI_Request* request)
 {
-  if((*request) != MPI_REQUEST_NULL){
-    (*request)->refcount_--;
-    if((*request)->refcount_ < 0) {
-      (*request)->print_request("wrong refcount");
-      xbt_die("Whoops, wrong refcount");
-    }
-    if((*request)->refcount_==0){
-      if ((*request)->flags_ & MPI_REQ_GENERALIZED){
-        ((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state);
-        delete (*request)->generalized_funcs;
-      }else{
-        Comm::unref((*request)->comm_);
-        Datatype::unref((*request)->old_type_);
-      }
-      if ((*request)->op_!=MPI_REPLACE && (*request)->op_!=MPI_OP_NULL)
-        Op::unref(&(*request)->op_);
+  xbt_assert(*request != MPI_REQUEST_NULL, "freeing an already free request");
 
-      (*request)->print_request("Destroying");
-      delete *request;
-      *request = MPI_REQUEST_NULL;
-    }else{
-      (*request)->print_request("Decrementing");
+  (*request)->refcount_--;
+  if ((*request)->refcount_ < 0) {
+    (*request)->print_request("wrong refcount");
+    xbt_die("Whoops, wrong refcount");
+  }
+  if ((*request)->refcount_ == 0) {
+    if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
+      ((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state);
+    } else {
+      Comm::unref((*request)->comm_);
+      Datatype::unref((*request)->old_type_);
     }
-  }else{
-    xbt_die("freeing an already free request");
+    if ((*request)->op_ != MPI_REPLACE && (*request)->op_ != MPI_OP_NULL)
+      Op::unref(&(*request)->op_);
+
+    (*request)->print_request("Destroying");
+    F2C::free_f((*request)->f2c_id());
+    delete *request;
+    *request = MPI_REQUEST_NULL;
+  } else {
+    (*request)->print_request("Decrementing");
   }
 }
 
-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<MPI_Request>(a);
-  MPI_Request req = static_cast<MPI_Request>(b);
-  XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
-
-  xbt_assert(ref, "Cannot match recv against null reference");
-  xbt_assert(req, "Cannot match recv against null request");
-  if((ref->src_ == MPI_ANY_SOURCE || 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)
-      req->cancelled_=-1;//mark as uncancellable
+  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 %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());
+
+  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->flags_ & MPI_REQ_PROBE) == 0 ){
+      if (receiver->real_size_ < sender->real_size_){
+        XBT_DEBUG("Truncating message - should not happen: receiver size : %zu < sender size : %zu", receiver->real_size_, sender->real_size_);
+        receiver->truncated_ = true;
+      } else if (receiver->real_size_ > sender->real_size_){
+        receiver->real_size_=sender->real_size_;
+      }
+    }
+    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
+    req->flags_ |= MPI_REQ_MATCHED; // mark as impossible to cancel anymore
     XBT_DEBUG("match succeeded");
-    return 1;
-  }else return 0;
+    return true;
+  }
+  return false;
 }
 
-int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
+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 = 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;
+}
+
+bool Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*)
 {
-  MPI_Request ref = static_cast<MPI_Request>(a);
-  MPI_Request req = static_cast<MPI_Request>(b);
-  XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
-  xbt_assert(ref, "Cannot match send against null reference");
-  xbt_assert(req, "Cannot match send against null request");
-
-  if((req->src_ == MPI_ANY_SOURCE || 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 uncancellable
-    XBT_DEBUG("match succeeded");
-    return 1;
-  } else
-    return 0;
+  auto ref = static_cast<MPI_Request>(a);
+  auto req = static_cast<MPI_Request>(b);
+  return match_common(req, req, ref);
 }
 
-void Request::print_request(const char *message)
+bool Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
 {
-  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_);
+  auto ref = static_cast<MPI_Request>(a);
+  auto req = static_cast<MPI_Request>(b);
+  return match_common(req, ref, req);
 }
 
+void Request::print_request(const char* message) const
+{
+  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::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+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) : 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(),
-                     comm->group()->actor(dst)->get_pid(), 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(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+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(),
-                     comm->group()->actor(dst)->get_pid(), 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(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+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(),
-                     comm->group()->actor(dst)->get_pid(), 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(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
+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(),
-                          comm->group()->actor(dst)->get_pid(), 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(),
-                          comm->group()->actor(dst)->get_pid(), 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)
 {
+  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);
   return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
-                     src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
+                     source,
                      simgrid::s4u::this_actor::get_pid(), tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
 }
@@ -223,99 +240,150 @@ 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 */
+  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);
   if(op==MPI_OP_NULL){
-    request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
-                          comm->group()->actor(dst)->get_pid(), tag, comm,
+    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);
   }else{
-    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, 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)
 {
+  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);
   return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
-                     src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
-                     simgrid::s4u::this_actor::get_pid(), tag, comm,
+                     source, 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::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(),
-                        comm->group()->actor(dst)->get_pid(), 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();
+  return request;
+}
+
+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);
-  request->start();
+  if(dst != MPI_PROC_NULL)
+    request->start();
   return request;
 }
 
-MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+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(),
-                        comm->group()->actor(dst)->get_pid(), 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);
-  request->start();
+  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 */
+  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,
-                        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();
+                        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;
 }
 
-void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
+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);
-  wait(&request,status);
+  int retval = wait(&request,status);
   request = nullptr;
+  return retval;
 }
 
-void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+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(),
-                        comm->group()->actor(dst)->get_pid(), 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 | MPI_REQ_BSEND);
+
+  if(dst != MPI_PROC_NULL)
+   request->start();
+  wait(&request, MPI_STATUS_IGNORE);
+  request = nullptr;
+}
 
-  request->start();
+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);
+  if(dst != MPI_PROC_NULL)
+   request->start();
   wait(&request, MPI_STATUS_IGNORE);
   request = nullptr;
 }
 
-void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+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(),
-                        comm->group()->actor(dst)->get_pid(), 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);
 
-  request->start();
+  if(dst != MPI_PROC_NULL)
+   request->start();
   wait(&request,MPI_STATUS_IGNORE);
   request = nullptr;
 }
 
-void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
+void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
                        void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
                        MPI_Comm comm, MPI_Status * status)
 {
-  MPI_Request requests[2];
-  MPI_Status stats[2];
-  int myid = simgrid::s4u::this_actor::get_pid();
-  if ((comm->group()->actor(dst)->get_pid() == myid) && (comm->group()->actor(src)->get_pid() == myid)) {
+  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);
+  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;
+  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) {
-      status->MPI_SOURCE = src;
+      status->MPI_SOURCE = source;
       status->MPI_TAG    = recvtag;
       status->MPI_ERROR  = MPI_SUCCESS;
       status->count      = sendcount * sendtype->size();
@@ -324,8 +392,8 @@ void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int d
   }
   requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
   requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
-  startall(2, requests);
-  waitall(2, requests, stats);
+  startall(2, requests.data());
+  waitall(2, requests.data(), stats.data());
   unref(&requests[0]);
   unref(&requests[1]);
   if(status != MPI_STATUS_IGNORE) {
@@ -339,30 +407,35 @@ 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();
 
+  // 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_;
   if ((flags_ & MPI_REQ_RECV) != 0) {
     this->print_request("New recv");
 
     simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
 
-    int async_small_thresh = simgrid::config::get_value<int>("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<int>(size_) < async_small_thresh) {
+    } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(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<void*>(this));
+      simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
 
       if (action == nullptr) {
         mailbox = process->mailbox();
@@ -378,7 +451,7 @@ 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<void*>(this));
+      simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
 
       if (action == nullptr) {
         XBT_DEBUG("No, nothing in the permanent receive mailbox");
@@ -388,27 +461,24 @@ 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->get_actor()->get_impl(), mailbox->get_impl(), 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_ & 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_));
-    int rank = src_;
-    if (TRACE_smpi_view_internals()) {
-      TRACE_smpi_send(rank, rank, dst_, tag_, size_);
-    }
+    const simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(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");
 
     void* buf = buf_;
     if ((flags_ & MPI_REQ_SSEND) == 0 &&
-        ((flags_ & MPI_REQ_RMA) != 0 ||
-         static_cast<int>(size_) < simgrid::config::get_value<int>("smpi/send-is-detached-thresh"))) {
+        ((flags_ & MPI_REQ_RMA) != 0 || (flags_ & MPI_REQ_BSEND) != 0 ||
+         static_cast<int>(size_) < smpi_cfg_detached_send_thresh())) {
       void *oldbuf = nullptr;
       detached_    = true;
       XBT_DEBUG("Send request %p is detached", this);
@@ -416,12 +486,14 @@ 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<char*>(buf_) >= smpi_data_exe_start) &&
               (static_cast<char*>(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(simgrid::s4u::Actor::by_pid(src_));
           }
+          //we need this temporary buffer even for bsend, as it will be released in the copy callback and we don't have a way to differentiate it
+          //so actually ... don't use manually attached buffer space.
           buf = xbt_malloc(size_);
           memcpy(buf,oldbuf,size_);
           XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
@@ -439,23 +511,21 @@ void Request::start()
     }
 
     if(sleeptime > 0.0){
-      simcall_process_sleep(sleeptime);
+      simgrid::s4u::this_actor::sleep_for(sleeptime);
       XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
     }
 
-    int async_small_thresh = simgrid::config::get_value<int>("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<int>(size_) < async_small_thresh) { // eager mode
+    } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(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<void*>(this));
+      simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
       if (action == nullptr) {
         if ((flags_ & MPI_REQ_SSEND) == 0) {
           mailbox = process->mailbox_small();
@@ -479,12 +549,12 @@ void Request::start()
       XBT_DEBUG("Send request %p is in the large mailbox %s (buf: %p)", this, mailbox->get_cname(), buf_);
     }
 
-    // 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::by_pid(src_)->get_impl(), mailbox->get_impl(), size_, -1.0, buf, real_size_, &match_send,
+    size_t payload_size_ = size_ + 16;//MPI enveloppe size (tag+dest+communicator)
+    action_              = simcall_comm_isend(
+        simgrid::kernel::actor::ActorImpl::by_pid(src_), 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
-        not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this,
+        process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this,
         // detach if msg size < eager/rdv switch limit
         detached_);
     XBT_DEBUG("send simcall posted");
@@ -495,7 +565,7 @@ 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();
   }
 }
@@ -506,75 +576,62 @@ void Request::startall(int count, MPI_Request * requests)
     return;
 
   for(int i = 0; i < count; i++) {
-    requests[i]->start();
+    if(requests[i]->src_ != MPI_PROC_NULL && requests[i]->dst_ != MPI_PROC_NULL)
+      requests[i]->start();
   }
 }
 
 void Request::cancel()
 {
-  if(cancelled_!=-1)
-    cancelled_=1;
+  this->flags_ |= MPI_REQ_CANCELLED;
   if (this->action_ != nullptr)
     (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
 }
 
 int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
-  //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
+  // 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
   //     while (MPI_Test(request, flag, status) && flag) dostuff...
   // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
   // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
+  xbt_assert(*request != MPI_REQUEST_NULL);
+
   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)
-    simcall_process_sleep(nsleeps*smpi_test_sleep);
+    simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
 
   Status::empty(status);
   *flag = 1;
-  if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) {
-    if ((*request)->action_ != nullptr){
+  if (((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) == 0) {
+    if ((*request)->action_ != nullptr && ((*request)->flags_ & MPI_REQ_CANCELLED) == 0){
       try{
-        *flag = simcall_comm_test((*request)->action_);
-      }catch (xbt_ex& e) {
+        *flag = simcall_comm_test((*request)->action_.get());
+      } catch (const Exception&) {
         *flag = 0;
         return ret;
       }
     }
-    if (*request != MPI_REQUEST_NULL && 
-        ((*request)->flags_ & MPI_REQ_GENERALIZED)
-        && !((*request)->flags_ & MPI_REQ_COMPLETE)) 
+    if (((*request)->flags_ & MPI_REQ_GENERALIZED) && !((*request)->flags_ & MPI_REQ_COMPLETE))
       *flag=0;
     if (*flag) {
-      finish_wait(request,status);
+      finish_wait(request, status); // may invalidate *request
       if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
+        MPI_Status tmp_status;
         MPI_Status* mystatus;
-        if(status==MPI_STATUS_IGNORE){
-          mystatus=new MPI_Status();
+        if (status == MPI_STATUS_IGNORE) {
+          mystatus = &tmp_status;
           Status::empty(mystatus);
-        }else{
-          mystatus=status;
+        } else {
+          mystatus = status;
         }
         ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
-        if(status==MPI_STATUS_IGNORE) 
-          delete mystatus;
       }
       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<bool>("smpi/grow-injected-times")) {
+    } else if (smpi_cfg_grow_injected_times()) {
       nsleeps++;
     }
   }
@@ -583,7 +640,6 @@ int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
 
 int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[])
 {
-  int ret = MPI_SUCCESS;
   int error=0;
   int count_dead = 0;
   int flag = 0;
@@ -592,15 +648,15 @@ int Request::testsome(int incount, MPI_Request requests[], int *count, int *indi
 
   *count = 0;
   for (int i = 0; i < incount; i++) {
-    if (requests[i] != MPI_REQUEST_NULL) {
-      ret = test(&requests[i], pstat, &flag);
+    if (requests[i] != MPI_REQUEST_NULL && not (requests[i]->flags_ & MPI_REQ_FINISHED)) {
+      int ret = test(&requests[i], pstat, &flag);
       if(ret!=MPI_SUCCESS)
         error = 1;
       if(flag) {
-        indices[i] = 1;
-        (*count)++;
+        indices[*count] = i;
         if (status != MPI_STATUSES_IGNORE)
-          status[i] = *pstat;
+          status[*count] = *pstat;
+        (*count)++;
         if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
           requests[i] = MPI_REQUEST_NULL;
       }
@@ -636,10 +692,11 @@ int Request::testany(int count, MPI_Request requests[], int *index, int* flag, M
     //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
     static int nsleeps = 1;
     if(smpi_test_sleep > 0)
-      simcall_process_sleep(nsleeps*smpi_test_sleep);
+      simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
     try{
       i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
-    }catch (xbt_ex& e) {
+    } catch (const Exception&) {
+      XBT_DEBUG("Exception in testany");
       return 0;
     }
     
@@ -652,20 +709,20 @@ int Request::testany(int count, MPI_Request requests[], int *index, int* flag, M
       } else {
         finish_wait(&requests[*index],status);
       if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){
+        MPI_Status tmp_status;
         MPI_Status* mystatus;
-        if(status==MPI_STATUS_IGNORE){
-          mystatus=new MPI_Status();
+        if (status == MPI_STATUS_IGNORE) {
+          mystatus = &tmp_status;
           Status::empty(mystatus);
-        }else{
-          mystatus=status;
+        } else {
+          mystatus = status;
         }
         ret=(requests[*index]->generalized_funcs)->query_fn((requests[*index]->generalized_funcs)->extra_state, mystatus);
-        if(status==MPI_STATUS_IGNORE) 
-          delete mystatus;
       }
 
         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;
       }
       nsleeps = 1;
@@ -673,8 +730,10 @@ int Request::testany(int count, MPI_Request requests[], int *index, int* flag, M
       nsleeps++;
     }
   } else {
+      XBT_DEBUG("Testany on inactive handles, returning flag=1 but empty status");
       //all requests are null or inactive, return true
       *flag = 1;
+      *index = MPI_UNDEFINED;
       Status::empty(status);
   }
 
@@ -687,11 +746,10 @@ int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status
   MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
   int flag;
   int error = 0;
-  int ret=MPI_SUCCESS;
   *outflag = 1;
   for(int i=0; i<count; i++){
     if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
-      ret = test(&requests[i], pstat, &flag);
+      int ret = test(&requests[i], pstat, &flag);
       if (flag){
         flag=0;
         requests[i]=MPI_REQUEST_NULL;
@@ -730,10 +788,10 @@ 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<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)->get_pid(),
-                                    simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV);
+  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),
+                  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 *
@@ -752,8 +810,8 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   s4u::Mailbox* 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 (simgrid::config::get_value<int>("smpi/async-small-thresh") > 0) {
+  // We have to test both mailboxes as we don't know if we will receive one or another
+  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<void*>(request));
@@ -767,7 +825,7 @@ 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<kernel::activity::CommImpl>(request->action_);
-    MPI_Request req                         = static_cast<MPI_Request>(sync_comm->src_data_);
+    const Request* req                      = static_cast<MPI_Request>(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_);
@@ -779,19 +837,42 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   }
   else {
     *flag = 0;
-    if (simgrid::config::get_value<bool>("smpi/grow-injected-times"))
+    if (smpi_cfg_grow_injected_times())
       nsleeps++;
   }
   unref(&request);
   xbt_assert(request == MPI_REQUEST_NULL);
 }
 
+int Request::finish_nbc_requests(MPI_Request* request){
+  int ret = waitall((*request)->nbc_requests_size_, (*request)->nbc_requests_, MPI_STATUSES_IGNORE);
+  XBT_DEBUG("finish non blocking collective request with %d sub-requests", (*request)->nbc_requests_size_);
+  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;
+  return ret;
+}
+
 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
 {
   MPI_Request req = *request;
   Status::empty(status);
-  
-  if (req->cancelled_==1){
+  if((req->flags_ & MPI_REQ_CANCELLED) != 0 && (req->flags_ & MPI_REQ_MATCHED) == 0) {
     if (status!=MPI_STATUS_IGNORE)
       status->cancelled=1;
     if(req->detached_sender_ != nullptr)
@@ -800,57 +881,70 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
     return;
   }
 
-  if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0)) && ((req->flags_ & MPI_REQ_PREPARED) == 0) &&
-      ((req->flags_ & MPI_REQ_GENERALIZED) == 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);
-      status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
-      status->MPI_ERROR  = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
+  if ((*request)->flags() & MPI_REQ_NBC){
+    int ret = finish_nbc_requests(request);
+    if (ret != MPI_SUCCESS)
+      xbt_die("error when finishing non blocking collective requests");
+  }
+
+  if ((req->flags_ & (MPI_REQ_PREPARED | MPI_REQ_GENERALIZED | MPI_REQ_FINISHED)) == 0) {
+    if (status != MPI_STATUS_IGNORE) {
+      if (req->src_== MPI_PROC_NULL || req->dst_== MPI_PROC_NULL){
+        Status::empty(status);
+        status->MPI_SOURCE = MPI_PROC_NULL;
+      } else {
+        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;
+      }
       // this handles the case were size in receive differs from size in send
       status->count = req->real_size_;
     }
+    //detached send will be finished at the other end
+    if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0))) {
+      req->print_request("Finishing");
+      MPI_Datatype datatype = req->old_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_cfg_privatization() != SmpiPrivStrategies::NONE &&
+            static_cast<char*>(req->old_buf_) >= smpi_data_exe_start &&
+            static_cast<char*>(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());
+        }
 
-    req->print_request("Finishing");
-    MPI_Datatype datatype = req->old_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_privatize_global_variables != SmpiPrivStrategies::NONE &&
-          static_cast<char*>(req->old_buf_) >= smpi_data_exe_start &&
-          static_cast<char*>(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());
-      }
-
-      if(datatype->flags() & DT_FLAG_DERIVED){
-        // This part handles the problem of non-contignous memory the unserialization at the reception
-        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_ & 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);
+        if(datatype->flags() & DT_FLAG_DERIVED){
+          // This part handles the problem of non-contiguous memory the unserialization at the reception
+          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;
         }
-        xbt_free(req->buf_);
       }
     }
   }
 
   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){
     //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());
-    if(sleeptime > 0.0){
-      simcall_process_sleep(sleeptime);
+    if (sleeptime > 0.0) {
+      simgrid::s4u::this_actor::sleep_for(sleeptime);
       XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
     }
     unref(&(req->detached_sender_));
@@ -858,41 +952,44 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
   if (req->flags_ & MPI_REQ_PERSISTENT)
     req->action_ = nullptr;
   req->flags_ |= MPI_REQ_FINISHED;
-  unref(request);
+
+  if (req->truncated_) {
+    char error_string[MPI_MAX_ERROR_STRING];
+    int error_size;
+    PMPI_Error_string(MPI_ERR_TRUNCATE, error_string, &error_size);
+    MPI_Errhandler err = (req->comm_) ? (req->comm_)->errhandler() : MPI_ERRHANDLER_NULL;
+    if (err == MPI_ERRHANDLER_NULL || err == MPI_ERRORS_RETURN)
+      XBT_WARN("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
+    else if (err == MPI_ERRORS_ARE_FATAL)
+      xbt_die("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
+    else
+      err->call((req->comm_), MPI_ERR_TRUNCATE);
+    if (err != MPI_ERRHANDLER_NULL)
+      simgrid::smpi::Errhandler::unref(err);
+    MC_assert(not MC_is_active()); /* Only fail in MC mode */
+  }
+  if(req->src_ != MPI_PROC_NULL && req->dst_ != MPI_PROC_NULL)
+    unref(request);
 }
 
 int Request::wait(MPI_Request * request, MPI_Status * status)
 {
+  // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Wait before)
+  xbt_assert(*request != MPI_REQUEST_NULL);
+
   int ret=MPI_SUCCESS;
-  // 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(buf);
-        }
-      }
-      if((*request)->nbc_requests_[i]!=MPI_REQUEST_NULL)
-        Request::unref(&((*request)->nbc_requests_[i]));
+
+  if((*request)->src_ == MPI_PROC_NULL || (*request)->dst_ == MPI_PROC_NULL){
+    if (status != MPI_STATUS_IGNORE) {
+      Status::empty(status);
+      status->MPI_SOURCE = MPI_PROC_NULL;
     }
-    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) {
+  if ((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) {
     Status::empty(status);
     return ret;
   }
@@ -900,31 +997,33 @@ 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);
-      }catch (xbt_ex& e) {
+        simcall_comm_wait((*request)->action_.get(), -1.0);
+      } catch (const Exception&) {
         XBT_VERB("Request cancelled");
       }
   }
 
-  if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
-    MPI_Status* mystatus;
+  if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
     if(!((*request)->flags_ & MPI_REQ_COMPLETE)){
       ((*request)->generalized_funcs)->mutex->lock();
       ((*request)->generalized_funcs)->cond->wait(((*request)->generalized_funcs)->mutex);
       ((*request)->generalized_funcs)->mutex->unlock();
-      }
-    if(status==MPI_STATUS_IGNORE){
-      mystatus=new MPI_Status();
+    }
+    MPI_Status tmp_status;
+    MPI_Status* mystatus;
+    if (status == MPI_STATUS_IGNORE) {
+      mystatus = &tmp_status;
       Status::empty(mystatus);
-    }else{
-      mystatus=status;
+    } else {
+      mystatus = status;
     }
     ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
-    if(status==MPI_STATUS_IGNORE) 
-      delete mystatus;
   }
 
-  finish_wait(request,status);
+  if ((*request)->truncated_)
+    ret = MPI_ERR_TRUNCATE;
+
+  finish_wait(request, status); // may invalidate *request
   if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
     *request = MPI_REQUEST_NULL;
   return ret;
@@ -932,12 +1031,11 @@ int Request::wait(MPI_Request * request, MPI_Status * status)
 
 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
 {
-  std::vector<simgrid::kernel::activity::CommImpl*> comms;
-  comms.reserve(count);
   int index = MPI_UNDEFINED;
 
   if(count > 0) {
     // Wait for a request to complete
+    std::vector<simgrid::kernel::activity::CommImpl*> comms;
     std::vector<int> map;
     XBT_DEBUG("Wait for one of %d", count);
     for(int i = 0; i < count; i++) {
@@ -949,7 +1047,7 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
           map.push_back(i);
         } else {
           // This is a finished detached request, let's return this one
-          comms.clear(); // so we free don't do the waitany call
+          comms.clear(); // don't do the waitany call afterwards
           index = i;
           finish_wait(&requests[i], status); // cleanup if refcount = 0
           if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
@@ -960,13 +1058,12 @@ 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=MPI_UNDEFINED;
+      int i;
       try{
-        // this is not a detached send
         i = simcall_comm_waitany(comms.data(), comms.size(), -1);
-      }catch (xbt_ex& e) {
-      XBT_INFO("request %d cancelled ",i);
-        return i;
+      } catch (const Exception&) {
+        XBT_INFO("request cancelled");
+        i = -1;
       }
 
       // not MPI_UNDEFINED, as this is a simix return code
@@ -989,7 +1086,7 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
   return index;
 }
 
-static int sort_accumulates(MPI_Request a, MPI_Request b)
+static int sort_accumulates(const Request* a, const Request* b)
 {
   return (a->tag() > b->tag());
 }
@@ -1018,8 +1115,8 @@ int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
       wait(&requests[c],pstat);
       index = c;
     } else {
-      index = waitany(count, (MPI_Request*)requests, pstat);
-      
+      index = waitany(count, requests, pstat);
+
       if (index == MPI_UNDEFINED)
         break;
 
@@ -1053,8 +1150,7 @@ int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Sta
   int index = 0;
   MPI_Status stat;
   MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
-
-  index = waitany(incount, (MPI_Request*)requests, pstat);
+  index             = waitany(incount, requests, pstat);
   if(index==MPI_UNDEFINED) return MPI_UNDEFINED;
   if(status != MPI_STATUSES_IGNORE) {
     status[count] = *pstat;
@@ -1062,7 +1158,8 @@ int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Sta
   indices[count] = index;
   count++;
   for (int i = 0; i < incount; i++) {
-    if (requests[i] != MPI_REQUEST_NULL) {
+    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;
@@ -1078,38 +1175,26 @@ int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Sta
   return count;
 }
 
-MPI_Request Request::f2c(int id) {
-  char key[KEY_SIZE];
-  if(id==MPI_FORTRAN_REQUEST_NULL)
-    return static_cast<MPI_Request>(MPI_REQUEST_NULL);
-  return static_cast<MPI_Request>(F2C::f2c_lookup()->at(get_key_id(key, id)));
-}
-
-int Request::add_f()
+MPI_Request Request::f2c(int id)
 {
-  if (F2C::f2c_lookup() == nullptr) {
-    F2C::set_f2c_lookup(new std::unordered_map<std::string, F2C*>);
-  }
-  char key[KEY_SIZE];
-  (*(F2C::f2c_lookup()))[get_key_id(key, F2C::f2c_id())] = this;
-  F2C::f2c_id_increment();
-  return F2C::f2c_id()-1;
+  if(id==MPI_FORTRAN_REQUEST_NULL)
+    return MPI_REQUEST_NULL;
+  return static_cast<MPI_Request>(F2C::lookup()->at(id));
 }
 
 void Request::free_f(int id)
 {
   if (id != MPI_FORTRAN_REQUEST_NULL) {
-    char key[KEY_SIZE];
-    F2C::f2c_lookup()->erase(get_key_id(key, id));
+    F2C::lookup()->erase(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) {
-    req->iprobe(req->src_, req->tag_, req->comm_, flag, status);
+    req->iprobe(req->comm_->group()->rank(req->src_), req->tag_, req->comm_, flag, status);
     if(*flag)
       return MPI_SUCCESS;
   }
@@ -1123,7 +1208,7 @@ int Request::get_status(MPI_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;
@@ -1132,13 +1217,14 @@ 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;
   (*request)->refcount_ = 1;
-  ((*request)->generalized_funcs) = new s_smpi_mpi_generalized_request_funcs_t;
+  ((*request)->generalized_funcs)             = std::make_unique<smpi_mpi_generalized_request_funcs_t>();
   ((*request)->generalized_funcs)->query_fn=query_fn;
   ((*request)->generalized_funcs)->free_fn=free_fn;
   ((*request)->generalized_funcs)->cancel_fn=cancel_fn;
@@ -1148,8 +1234,9 @@ int Request::grequest_start( MPI_Grequest_query_function *query_fn, MPI_Grequest
   return MPI_SUCCESS;
 }
 
-int Request::grequest_complete( MPI_Request request){
-  if ((!(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex==NULL) 
+int Request::grequest_complete(MPI_Request request)
+{
+  if ((!(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex == nullptr)
     return MPI_ERR_REQUEST;
   request->generalized_funcs->mutex->lock();
   request->flags_ |= MPI_REQ_COMPLETE; // in case wait would be called after complete
@@ -1168,13 +1255,14 @@ void Request::set_nbc_requests(MPI_Request* reqs, int size){
   }
 }
 
-int Request::get_nbc_requests_size(){
+int Request::get_nbc_requests_size() const
+{
   return nbc_requests_size_;
 }
 
-MPI_Request* Request::get_nbc_requests(){
+MPI_Request* Request::get_nbc_requests() const
+{
   return nbc_requests_;
 }
-
 }
 }