X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/cba7a4411e9eb8cd10066db126c413363c65111a..061f1e9720921cc227dab4f6ddeca9d30028e5d4:/src/smpi/mpi/smpi_request.cpp

diff --git a/src/smpi/mpi/smpi_request.cpp b/src/smpi/mpi/smpi_request.cpp
index 64c35749a2..076207389f 100644
--- a/src/smpi/mpi/smpi_request.cpp
+++ b/src/smpi/mpi/smpi_request.cpp
@@ -1,23 +1,28 @@
-/* Copyright (c) 2007-2017. The SimGrid Team. All rights reserved.          */
+/* Copyright (c) 2007-2019. The SimGrid Team. All rights reserved.          */
 
 /* This program is free software; you can redistribute it and/or modify it
  * under the terms of the license (GNU LGPL) which comes with this package. */
-
+#include "simgrid/s4u/Mutex.hpp"
+#include "simgrid/s4u/ConditionVariable.hpp"
 #include "smpi_request.hpp"
 
-#include "SmpiHost.hpp"
 #include "mc/mc.h"
 #include "private.hpp"
+#include "simgrid/Exception.hpp"
+#include "simgrid/s4u/Exec.hpp"
 #include "smpi_comm.hpp"
 #include "smpi_datatype.hpp"
+#include "smpi_host.hpp"
 #include "smpi_op.hpp"
-#include "smpi_process.hpp"
 #include "src/kernel/activity/CommImpl.hpp"
 #include "src/mc/mc_replay.hpp"
+#include "src/smpi/include/smpi_actor.hpp"
+#include "xbt/config.hpp"
+
 
 #include <algorithm>
 
-XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (reques)");
+XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (request)");
 
 static simgrid::config::Flag<double> smpi_iprobe_sleep(
   "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
@@ -26,24 +31,24 @@ static simgrid::config::Flag<double> smpi_test_sleep(
 
 std::vector<s_smpi_factor_t> smpi_ois_values;
 
-extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
+extern void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*, size_t);
 
 namespace simgrid{
 namespace smpi{
 
-Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags)
-    : buf_(buf), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags)
+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)
 {
   void *old_buf = nullptr;
 // FIXME Handle the case of a partial shared malloc.
-  if ((((flags & RECV) != 0) && ((flags & ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
+  if ((((flags & MPI_REQ_RECV) != 0) && ((flags & MPI_REQ_ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
     // This part handles the problem of non-contiguous memory
     old_buf = buf;
     if (count==0){
       buf_ = nullptr;
     }else {
       buf_ = xbt_malloc(count*datatype->size());
-      if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
+      if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & MPI_REQ_SEND) != 0)) {
         datatype->serialize(old_buf, buf_, count);
       }
     }
@@ -53,6 +58,8 @@ Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst,
   size_ = datatype->size() * count;
   datatype->ref();
   comm_->ref();
+  if(op != MPI_REPLACE && op != MPI_OP_NULL)
+    op_->ref();
   action_          = nullptr;
   detached_        = 0;
   detached_sender_ = nullptr;
@@ -60,56 +67,41 @@ Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst,
   truncated_       = 0;
   real_size_       = 0;
   real_tag_        = 0;
-  if (flags & PERSISTENT)
+  if (flags & MPI_REQ_PERSISTENT)
     refcount_ = 1;
   else
     refcount_ = 0;
-  op_   = MPI_REPLACE;
-}
-
-MPI_Comm Request::comm(){
-  return comm_;
+  cancelled_ = 0;
+  generalized_funcs=nullptr;
+  nbc_requests_=nullptr;
+  nbc_requests_size_=0;
 }
 
-int Request::src(){
-  return src_;
-}
-
-int Request::dst(){
-  return dst_;
-}
-
-int Request::tag(){
-  return tag_;
-}
-
-int Request::flags(){
-  return flags_;
-}
-
-int Request::detached(){
-  return detached_;
-}
-
-size_t Request::size(){
-  return size_;
-}
-
-size_t Request::real_size(){
-  return real_size_;
+void Request::ref(){
+  refcount_++;
 }
 
 void Request::unref(MPI_Request* request)
 {
   if((*request) != MPI_REQUEST_NULL){
     (*request)->refcount_--;
-    if((*request)->refcount_<0) xbt_die("wrong refcount");
+    if((*request)->refcount_ < 0) {
+      (*request)->print_request("wrong refcount");
+      xbt_die("Whoops, wrong refcount");
+    }
     if((*request)->refcount_==0){
-        Datatype::unref((*request)->old_type_);
+      if ((*request)->flags_ & MPI_REQ_GENERALIZED){
+        ((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state);
+      }else{
         Comm::unref((*request)->comm_);
-        (*request)->print_request("Destroying");
-        delete *request;
-        *request = MPI_REQUEST_NULL;
+        Datatype::unref((*request)->old_type_);
+      }
+      if ((*request)->op_!=MPI_REPLACE && (*request)->op_!=MPI_OP_NULL)
+        Op::unref(&(*request)->op_);
+
+      (*request)->print_request("Destroying");
+      delete *request;
+      *request = MPI_REQUEST_NULL;
     }else{
       (*request)->print_request("Decrementing");
     }
@@ -118,7 +110,7 @@ void Request::unref(MPI_Request* request)
   }
 }
 
-int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
+int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*)
 {
   MPI_Request ref = static_cast<MPI_Request>(a);
   MPI_Request req = static_cast<MPI_Request>(b);
@@ -137,12 +129,14 @@ int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* i
       ref->truncated_ = 1;
     if(req->detached_==1)
       ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
+    if(req->cancelled_==0)
+      req->cancelled_=-1;//mark as uncancellable
     XBT_DEBUG("match succeeded");
     return 1;
   }else return 0;
 }
 
-int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
+int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
 {
   MPI_Request ref = static_cast<MPI_Request>(a);
   MPI_Request req = static_cast<MPI_Request>(b);
@@ -160,6 +154,8 @@ int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl* i
       req->truncated_ = 1;
     if(ref->detached_==1)
       req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
+    if(req->cancelled_==0)
+      req->cancelled_=-1;//mark as uncancellable
     XBT_DEBUG("match succeeded");
     return 1;
   } else
@@ -177,20 +173,23 @@ void Request::print_request(const char *message)
 MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
 {
 
-  return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
-                          comm->group()->index(dst), tag, comm, PERSISTENT | SEND | PREPARED);
+  return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                     comm->group()->actor(dst)->get_pid(), tag, comm,
+                     MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
 }
 
 MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
 {
-  return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
-                        comm->group()->index(dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
+  return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                     comm->group()->actor(dst)->get_pid(), tag, comm,
+                     MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
 }
 
 MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
 {
-  return new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, smpi_process()->index(),
-                          comm->group()->index(dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
+  return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                     comm->group()->actor(dst)->get_pid(), tag, comm,
+                     MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
 }
 
 
@@ -199,21 +198,24 @@ MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype,
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   if(op==MPI_OP_NULL){
-    request = new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag,
-                            comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
+    request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
+                          comm->group()->actor(dst)->get_pid(), tag, comm,
+                          MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
   }else{
-    request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,  src, dst, tag,
-                            comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
-    request->op_ = op;
+    request      = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
+                          comm->group()->actor(dst)->get_pid(), tag, comm,
+                          MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
+                              MPI_REQ_ACCUMULATE, op);
   }
   return request;
 }
 
 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
 {
-  return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,
-                          src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src),
-                          smpi_process()->index(), tag, comm, PERSISTENT | RECV | PREPARED);
+  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,
+                     MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
 }
 
 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
@@ -221,12 +223,13 @@ MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype,
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   if(op==MPI_OP_NULL){
-    request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag, comm,
-                          RMA | NON_PERSISTENT | RECV | PREPARED);
+    request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
+                          comm->group()->actor(dst)->get_pid(), tag, comm,
+                          MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
   }else{
-    request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag, comm,
-                          RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
-    request->op_ = op;
+    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);
   }
   return request;
 }
@@ -234,15 +237,17 @@ MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype,
 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
 {
   return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
-                     src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src), smpi_process()->index(), tag,
-                     comm, PERSISTENT | RECV | PREPARED);
+                     src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
+                     simgrid::s4u::this_actor::get_pid(), tag, comm,
+                     MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
 }
 
 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
-  request             = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
-                        comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
+  request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                        comm->group()->actor(dst)->get_pid(), tag, comm,
+                        MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
   request->start();
   return request;
 }
@@ -250,8 +255,9 @@ MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst,
 MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
-  request             = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
-                        comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SSEND | SEND);
+  request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                        comm->group()->actor(dst)->get_pid(), tag, comm,
+                        MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
   request->start();
   return request;
 }
@@ -261,8 +267,8 @@ MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src,
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   request             = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
-                        src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src), smpi_process()->index(),
-                        tag, comm, NON_PERSISTENT | RECV);
+                        src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
+                        simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
   request->start();
   return request;
 }
@@ -278,8 +284,8 @@ void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag
 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
-  request             = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
-                        comm->group()->index(dst), tag, comm, NON_PERSISTENT | SEND);
+  request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                        comm->group()->actor(dst)->get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
 
   request->start();
   wait(&request, MPI_STATUS_IGNORE);
@@ -289,8 +295,9 @@ void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag
 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
-  request             = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
-                        comm->group()->index(dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
+  request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+                        comm->group()->actor(dst)->get_pid(), tag, comm,
+                        MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
 
   request->start();
   wait(&request,MPI_STATUS_IGNORE);
@@ -303,10 +310,16 @@ void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int d
 {
   MPI_Request requests[2];
   MPI_Status stats[2];
-  int myid=smpi_process()->index();
-  if ((comm->group()->index(dst) == myid) && (comm->group()->index(src) == myid)){
-      Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
-      return;
+  int myid = simgrid::s4u::this_actor::get_pid();
+  if ((comm->group()->actor(dst)->get_pid() == myid) && (comm->group()->actor(src)->get_pid() == myid)) {
+    Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
+    if (status != MPI_STATUS_IGNORE) {
+      status->MPI_SOURCE = src;
+      status->MPI_TAG    = recvtag;
+      status->MPI_ERROR  = MPI_SUCCESS;
+      status->count      = sendcount * sendtype->size();
+    }
+    return;
   }
   requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
   requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
@@ -322,40 +335,40 @@ void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int d
 
 void Request::start()
 {
-  smx_mailbox_t mailbox;
+  s4u::Mailbox* mailbox;
 
   xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
-  flags_ &= ~PREPARED;
-  flags_ &= ~FINISHED;
-  refcount_++;
+  flags_ &= ~MPI_REQ_PREPARED;
+  flags_ &= ~MPI_REQ_FINISHED;
+  this->ref();
 
-  if ((flags_ & RECV) != 0) {
+  if ((flags_ & MPI_REQ_RECV) != 0) {
     this->print_request("New recv");
 
-    simgrid::smpi::Process* process = smpi_process_remote(dst_);
+    simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
 
-    int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
+    int async_small_thresh = simgrid::config::get_value<int>("smpi/async-small-thresh");
 
-    xbt_mutex_t mut = process->mailboxes_mutex();
-    if (async_small_thresh != 0 || (flags_ & RMA) != 0)
-      xbt_mutex_acquire(mut);
+    simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
+    if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
+      mut->lock();
 
-    if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) {
+    if (async_small_thresh == 0 && (flags_ & MPI_REQ_RMA) == 0) {
       mailbox = process->mailbox();
-    }
-    else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
+    } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
       //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
       //begin with the more appropriate one : the small one.
       mailbox = process->mailbox_small();
-      XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
-      smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
+      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));
 
       if (action == nullptr) {
         mailbox = process->mailbox();
-        XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
-        action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
+        XBT_DEBUG("No, nothing in the small mailbox test the other one : %s", mailbox->get_cname());
+        action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
         if (action == nullptr) {
-          XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
+          XBT_DEBUG("Still nothing, switch back to the small mailbox : %s", mailbox->get_cname());
           mailbox = process->mailbox_small();
         }
       } else {
@@ -364,7 +377,7 @@ void Request::start()
     } else {
       mailbox = process->mailbox_small();
       XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
-      smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
+      smx_activity_t action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
 
       if (action == nullptr) {
         XBT_DEBUG("No, nothing in the permanent receive mailbox");
@@ -377,14 +390,14 @@ void Request::start()
     // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
     real_size_=size_;
     action_   = simcall_comm_irecv(
-        process->process(), mailbox, buf_, &real_size_, &match_recv,
+        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_ & RMA) != 0 )
-      xbt_mutex_release(mut);
+    if (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::Process* process = smpi_process_remote(dst_);
+    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_);
@@ -392,19 +405,21 @@ void Request::start()
     this->print_request("New send");
 
     void* buf = buf_;
-    if ((flags_ & SSEND) == 0 && ( (flags_ & RMA) != 0
-        || static_cast<int>(size_) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) {
+    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"))) {
       void *oldbuf = nullptr;
       detached_ = 1;
       XBT_DEBUG("Send request %p is detached", this);
-      refcount_++;
+      this->ref();
       if (not(old_type_->flags() & DT_FLAG_DERIVED)) {
         oldbuf = buf_;
         if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
-          if ((smpi_privatize_global_variables != 0) && (static_cast<char*>(buf_) >= smpi_data_exe_start) &&
+          if ((smpi_privatize_global_variables != 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(src_);
+            smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_));
           }
           buf = xbt_malloc(size_);
           memcpy(buf,oldbuf,size_);
@@ -415,11 +430,11 @@ void Request::start()
 
     //if we are giving back the control to the user without waiting for completion, we have to inject timings
     double sleeptime = 0.0;
-    if (detached_ != 0 || ((flags_ & (ISEND | SSEND)) != 0)) { // issend should be treated as isend
+    if (detached_ != 0 || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
       // isend and send timings may be different
-      sleeptime = ((flags_ & ISEND) != 0)
-                      ? simgrid::s4u::Actor::self()->getHost()->extension<simgrid::smpi::SmpiHost>()->oisend(size_)
-                      : simgrid::s4u::Actor::self()->getHost()->extension<simgrid::smpi::SmpiHost>()->osend(size_);
+      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_);
     }
 
     if(sleeptime > 0.0){
@@ -427,27 +442,29 @@ void Request::start()
       XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
     }
 
-    int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
+    int async_small_thresh = simgrid::config::get_value<int>("smpi/async-small-thresh");
 
-    xbt_mutex_t mut=process->mailboxes_mutex();
+    simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
 
-    if (async_small_thresh != 0 || (flags_ & RMA) != 0)
-      xbt_mutex_acquire(mut);
+    if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
+      mut->lock();
 
-    if (not(async_small_thresh != 0 || (flags_ & RMA) != 0)) {
+    if (not(async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)) {
       mailbox = process->mailbox();
-    } else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
+    } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
       mailbox = process->mailbox();
-      XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
-      smx_activity_t action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
+      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));
       if (action == nullptr) {
-        if ((flags_ & SSEND) == 0){
+        if ((flags_ & MPI_REQ_SSEND) == 0) {
           mailbox = process->mailbox_small();
-          XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
+          XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %s",
+                    mailbox->get_cname());
         } else {
           mailbox = process->mailbox_small();
-          XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
-          action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
+          XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %s?",
+                    mailbox->get_cname());
+          action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
           if (action == nullptr) {
             XBT_DEBUG("No, we are first, send to large mailbox");
             mailbox = process->mailbox();
@@ -458,13 +475,13 @@ void Request::start()
       }
     } else {
       mailbox = process->mailbox();
-      XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
+      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(
-        SIMIX_process_from_PID(src_ + 1), mailbox, size_, -1.0, buf, real_size_, &match_send,
+        simgrid::s4u::Actor::by_pid(src_)->get_impl(), mailbox->get_impl(), size_, -1.0, buf, real_size_, &match_send,
         &xbt_free_f, // how to free the userdata if a detached send fails
         not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this,
         // detach if msg size < eager/rdv switch limit
@@ -472,10 +489,13 @@ void Request::start()
     XBT_DEBUG("send simcall posted");
 
     /* FIXME: detached sends are not traceable (action_ == nullptr) */
-    if (action_ != nullptr)
-      simcall_set_category(action_, TRACE_internal_smpi_get_category());
-    if (async_small_thresh != 0 || ((flags_ & RMA)!=0))
-      xbt_mutex_release(mut);
+    if (action_ != nullptr) {
+      std::string category = smpi_process()->get_tracing_category();
+      simgrid::simix::simcall([this, category] { this->action_->set_category(category); });
+    }
+
+    if (async_small_thresh != 0 || ((flags_ & MPI_REQ_RMA) != 0))
+      mut->unlock();
   }
 }
 
@@ -489,73 +509,126 @@ void Request::startall(int count, MPI_Request * requests)
   }
 }
 
-int Request::test(MPI_Request * request, MPI_Status * status) {
+void Request::cancel()
+{
+  if(cancelled_!=-1)
+    cancelled_=1;
+  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)
   // 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
   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);
 
+  MPI_Status* mystatus;
   Status::empty(status);
-  int flag = 1;
-  if (((*request)->flags_ & PREPARED) == 0) {
-    if ((*request)->action_ != nullptr)
-      flag = simcall_comm_test((*request)->action_);
-    if (flag) {
+  *flag = 1;
+  if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) {
+    if ((*request)->action_ != nullptr){
+      try{
+        *flag = simcall_comm_test((*request)->action_);
+      }catch (xbt_ex& e) {
+        *flag = 0;
+        return ret;
+      }
+    }
+    if (*request != MPI_REQUEST_NULL && 
+        ((*request)->flags_ & MPI_REQ_GENERALIZED)
+        && !((*request)->flags_ & MPI_REQ_COMPLETE)) 
+      *flag=0;
+    if (*flag) {
       finish_wait(request,status);
+      if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
+        if(status==MPI_STATUS_IGNORE){
+          mystatus=new MPI_Status();
+          Status::empty(mystatus);
+        }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_ & PERSISTENT) == 0)
+      if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0)
         *request = MPI_REQUEST_NULL;
-    } else if (xbt_cfg_get_boolean("smpi/grow-injected-times")){
+    } else if (simgrid::config::get_value<bool>("smpi/grow-injected-times")) {
       nsleeps++;
     }
   }
-  return flag;
+  return ret;
 }
 
-int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
+int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[])
 {
-  int count = 0;
+  int ret = MPI_SUCCESS;
+  int error=0;
   int count_dead = 0;
+  int flag = 0;
   MPI_Status stat;
   MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
 
+  *count = 0;
   for (int i = 0; i < incount; i++) {
     if (requests[i] != MPI_REQUEST_NULL) {
-      if (test(&requests[i], pstat)) {
+      ret = test(&requests[i], pstat, &flag);
+      if(ret!=MPI_SUCCESS)
+        error = 1;
+      if(flag) {
         indices[i] = 1;
-        count++;
+        (*count)++;
         if (status != MPI_STATUSES_IGNORE)
           status[i] = *pstat;
-        if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & NON_PERSISTENT))
+        if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
           requests[i] = MPI_REQUEST_NULL;
       }
     } else {
       count_dead++;
     }
   }
-  if(count_dead==incount)
-    return MPI_UNDEFINED;
-  else return count;
+  if(count_dead==incount)*count=MPI_UNDEFINED;
+  if(error!=0)
+    return MPI_ERR_IN_STATUS;
+  else
+    return MPI_SUCCESS;
 }
 
-int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
+int Request::testany(int count, MPI_Request requests[], int *index, int* flag, MPI_Status * status)
 {
-  std::vector<simgrid::kernel::activity::ActivityImplPtr> comms;
+  std::vector<simgrid::kernel::activity::CommImpl*> comms;
   comms.reserve(count);
 
   int i;
-  int flag = 0;
-
+  *flag = 0;
+  int ret = MPI_SUCCESS;
+  MPI_Status* mystatus;
   *index = MPI_UNDEFINED;
 
   std::vector<int> map; /** Maps all matching comms back to their location in requests **/
   for(i = 0; i < count; i++) {
-    if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & PREPARED)) {
-      comms.push_back(requests[i]->action_);
+    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);
     }
   }
@@ -564,40 +637,67 @@ int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status *
     static int nsleeps = 1;
     if(smpi_test_sleep > 0)
       simcall_process_sleep(nsleeps*smpi_test_sleep);
-
-    i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
+    try{
+      i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
+    }catch (xbt_ex& e) {
+      return 0;
+    }
+    
     if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
       *index = map[i];
-      finish_wait(&requests[*index],status);
-      flag             = 1;
-      nsleeps          = 1;
-      if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & NON_PERSISTENT)) {
-        requests[*index] = MPI_REQUEST_NULL;
+      if (requests[*index] != MPI_REQUEST_NULL && 
+          (requests[*index]->flags_ & MPI_REQ_GENERALIZED)
+          && !(requests[*index]->flags_ & MPI_REQ_COMPLETE)) {
+        *flag=0;
+      } else {
+        finish_wait(&requests[*index],status);
+      if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){
+        if(status==MPI_STATUS_IGNORE){
+          mystatus=new MPI_Status();
+          Status::empty(mystatus);
+        }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;
+        *flag=1;
       }
+      nsleeps = 1;
     } else {
       nsleeps++;
     }
   } else {
       //all requests are null or inactive, return true
-      flag = 1;
+      *flag = 1;
       Status::empty(status);
   }
 
-  return flag;
+  return ret;
 }
 
-int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
+int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status status[])
 {
   MPI_Status stat;
   MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
-  int flag=1;
+  int flag, 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_ & PREPARED)) {
-      if (test(&requests[i], pstat)!=1){
+    if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
+      ret = test(&requests[i], pstat, &flag);
+      if (flag){
         flag=0;
+        requests[i]=MPI_REQUEST_NULL;
       }else{
-          requests[i]=MPI_REQUEST_NULL;
+        *outflag=0;
       }
+      if (ret != MPI_SUCCESS) 
+        error = 1;
     }else{
       Status::empty(pstat);
     }
@@ -605,7 +705,10 @@ int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
       status[i] = *pstat;
     }
   }
-  return flag;
+  if(error==1) 
+    return MPI_ERR_IN_STATUS;
+  else 
+    return MPI_SUCCESS;
 }
 
 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
@@ -624,37 +727,47 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
   // This can speed up the execution of certain applications by an order of magnitude, such as HPL
   static int nsleeps = 1;
-  double speed        = simgrid::s4u::Actor::self()->getHost()->getSpeed();
-  double maxrate = xbt_cfg_get_double("smpi/iprobe-cpu-usage");
-  MPI_Request request = new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
-                 comm->group()->index(source), comm->rank(), tag, comm, PERSISTENT | RECV);
+  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);
   if (smpi_iprobe_sleep > 0) {
-    smx_activity_t iprobe_sleep = simcall_execution_start("iprobe", /* flops to executek*/nsleeps*smpi_iprobe_sleep*speed*maxrate, /* priority */1.0, /* performance bound */maxrate*speed);
-    simcall_execution_wait(iprobe_sleep);
+    /** Compute the number of flops we will sleep **/
+    s4u::this_actor::exec_init(/*nsleeps: See comment above */ nsleeps *
+                               /*(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. 
+         * 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)
+        ->start()
+        ->wait();
   }
   // behave like a receive, but don't do it
-  smx_mailbox_t mailbox;
+  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 (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
-      mailbox = smpi_process()->mailbox_small();
-      XBT_DEBUG("Trying to probe the perm recv mailbox");
-      request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
+  if (simgrid::config::get_value<int>("smpi/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));
   }
 
   if (request->action_ == nullptr){
     mailbox = smpi_process()->mailbox();
     XBT_DEBUG("trying to probe the other mailbox");
-    request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
+    request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
   }
 
   if (request->action_ != nullptr){
-    simgrid::kernel::activity::CommImplPtr sync_comm =
-        boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(request->action_);
-    MPI_Request req                            = static_cast<MPI_Request>(sync_comm->src_data);
+    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_);
     *flag = 1;
-    if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) {
+    if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) {
       status->MPI_SOURCE = comm->group()->rank(req->src_);
       status->MPI_TAG    = req->tag_;
       status->MPI_ERROR  = MPI_SUCCESS;
@@ -664,7 +777,7 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   }
   else {
     *flag = 0;
-    if (xbt_cfg_get_boolean("smpi/grow-injected-times"))
+    if (simgrid::config::get_value<bool>("smpi/grow-injected-times"))
       nsleeps++;
   }
   unref(&request);
@@ -674,8 +787,19 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
 {
   MPI_Request req = *request;
   Status::empty(status);
+  
+  if (req->cancelled_==1){
+    if (status!=MPI_STATUS_IGNORE)
+      status->cancelled=1;
+    if(req->detached_sender_ != nullptr)
+      unref(&(req->detached_sender_));
+    unref(request);
+    return;
+  }
 
-  if (not((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)) {
+  if (not((req->detached_ != 0) && ((req->flags_ & MPI_REQ_SEND) != 0)) 
+  && ((req->flags_ & MPI_REQ_PREPARED) == 0)
+  && ((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);
@@ -689,121 +813,159 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
     MPI_Datatype datatype = req->old_type_;
 
 // FIXME Handle the case of a partial shared malloc.
-    if (((req->flags_ & ACCUMULATE) != 0) ||
+    if (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) ||
         (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
 
-      if (not smpi_process()->replaying() && smpi_privatize_global_variables != 0 &&
+      if (not smpi_process()->replaying() && smpi_privatize_global_variables != SmpiPrivStrategies::NONE &&
           static_cast<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(smpi_process()->index());
+        smpi_switch_data_segment(simgrid::s4u::Actor::self());
       }
 
       if(datatype->flags() & DT_FLAG_DERIVED){
         // This part handles the problem of non-contignous memory the unserialization at the reception
-        if((req->flags_ & RECV) && datatype->size()!=0)
+        if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
           datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
         xbt_free(req->buf_);
-      }else if(req->flags_ & RECV){//apply op on contiguous buffer for accumulate
-          if(datatype->size()!=0){
-            int n =req->real_size_/datatype->size();
-            req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
-          }
-          xbt_free(req->buf_);
+      } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
+        if (datatype->size() != 0) {
+          int n = req->real_size_ / datatype->size();
+          req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
+        }
+        xbt_free(req->buf_);
       }
     }
   }
 
-  if (TRACE_smpi_view_internals() && ((req->flags_ & RECV) != 0)){
-    int rank = smpi_process()->index();
+  if (TRACE_smpi_view_internals() && ((req->flags_ & MPI_REQ_RECV) != 0)) {
+    int rank       = simgrid::s4u::this_actor::get_pid();
     int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
     TRACE_smpi_recv(src_traced, rank,req->tag_);
   }
   if(req->detached_sender_ != nullptr){
     //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
     double sleeptime =
-        simgrid::s4u::Actor::self()->getHost()->extension<simgrid::smpi::SmpiHost>()->orecv(req->real_size());
+        simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(req->real_size());
     if(sleeptime > 0.0){
       simcall_process_sleep(sleeptime);
       XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
     }
     unref(&(req->detached_sender_));
   }
-  if(req->flags_ & PERSISTENT)
+  if (req->flags_ & MPI_REQ_PERSISTENT)
     req->action_ = nullptr;
-  req->flags_ |= FINISHED;
+  req->flags_ |= MPI_REQ_FINISHED;
   unref(request);
 }
 
-void Request::wait(MPI_Request * request, MPI_Status * status)
+int Request::wait(MPI_Request * request, MPI_Status * status)
 {
+  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)->nbc_requests_[i]!=MPI_REQUEST_NULL)
+        Request::unref(&((*request)->nbc_requests_[i]));
+    }
+    delete[] (*request)->nbc_requests_;
+    (*request)->nbc_requests_size_=0;
+    unref(request);
+    return ret;
+  }
+
   (*request)->print_request("Waiting");
-  if ((*request)->flags_ & PREPARED) {
+  if ((*request)->flags_ & MPI_REQ_PREPARED) {
     Status::empty(status);
-    return;
+    return ret;
+  }
+
+  if ((*request)->action_ != nullptr){
+      try{
+        // this is not a detached send
+        simcall_comm_wait((*request)->action_, -1.0);
+      }catch (xbt_ex& e) {
+        XBT_VERB("Request cancelled");
+      }
   }
 
-  if ((*request)->action_ != nullptr)
-    // this is not a detached send
-    simcall_comm_wait((*request)->action_, -1.0);
+  if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
+    MPI_Status* mystatus;
+    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();
+      Status::empty(mystatus);
+    }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 != MPI_REQUEST_NULL && (((*request)->flags_ & NON_PERSISTENT)!=0))
+  if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
     *request = MPI_REQUEST_NULL;
+  return ret;
 }
 
 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
 {
-  s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
+  std::vector<simgrid::kernel::activity::CommImpl*> comms;
+  comms.reserve(count);
   int index = MPI_UNDEFINED;
 
   if(count > 0) {
-    int size = 0;
     // Wait for a request to complete
-    xbt_dynar_init(&comms, sizeof(smx_activity_t), [](void*ptr){
-      intrusive_ptr_release(*(simgrid::kernel::activity::ActivityImpl**)ptr);
-    });
-    int *map = xbt_new(int, count);
+    std::vector<int> map;
     XBT_DEBUG("Wait for one of %d", count);
     for(int i = 0; i < count; i++) {
-      if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED) &&
-          not(requests[i]->flags_ & FINISHED)) {
+      if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED) &&
+          not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
         if (requests[i]->action_ != nullptr) {
           XBT_DEBUG("Waiting any %p ", requests[i]);
-          intrusive_ptr_add_ref(requests[i]->action_.get());
-          xbt_dynar_push_as(&comms, simgrid::kernel::activity::ActivityImpl*, requests[i]->action_.get());
-          map[size] = i;
-          size++;
+          comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
+          map.push_back(i);
         } else {
           // This is a finished detached request, let's return this one
-          size  = 0; // so we free the dynar but don't do the waitany call
+          comms.clear(); // so we free don't do the waitany call
           index = i;
           finish_wait(&requests[i], status); // cleanup if refcount = 0
-          if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
+          if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
             requests[i] = MPI_REQUEST_NULL; // set to null
           break;
         }
       }
     }
-    if (size > 0) {
-      XBT_DEBUG("Enter waitany for %lu comms", xbt_dynar_length(&comms));
-      int i = simcall_comm_waitany(&comms, -1);
+    if (not comms.empty()) {
+      XBT_DEBUG("Enter waitany for %zu comms", comms.size());
+      int i=MPI_UNDEFINED;
+      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;
+      }
 
       // not MPI_UNDEFINED, as this is a simix return code
       if (i != -1) {
         index = map[i];
         //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
         if ((requests[index] == MPI_REQUEST_NULL) ||
-            (not((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))) {
+            (not((requests[index]->flags_ & MPI_REQ_ACCUMULATE) && (requests[index]->flags_ & MPI_REQ_RECV)))) {
           finish_wait(&requests[index],status);
-          if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
+          if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
             requests[index] = MPI_REQUEST_NULL;
         }
       }
     }
-
-    xbt_dynar_free_data(&comms);
-    xbt_free(map);
   }
 
   if (index==MPI_UNDEFINED)
@@ -827,7 +989,8 @@ int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
   //tag invalid requests in the set
   if (status != MPI_STATUSES_IGNORE) {
     for (int c = 0; c < count; c++) {
-      if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || (requests[c]->flags_ & PREPARED)) {
+      if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL ||
+          (requests[c]->flags_ & MPI_REQ_PREPARED)) {
         Status::empty(&status[c]);
       } else if (requests[c]->src_ == MPI_PROC_NULL) {
         Status::empty(&status[c]);
@@ -841,14 +1004,14 @@ int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
       index = c;
     } else {
       index = waitany(count, (MPI_Request*)requests, pstat);
+      
       if (index == MPI_UNDEFINED)
         break;
 
-      if (requests[index] != MPI_REQUEST_NULL
-           && (requests[index]->flags_ & RECV)
-           && (requests[index]->flags_ & ACCUMULATE))
+      if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_RECV) &&
+          (requests[index]->flags_ & MPI_REQ_ACCUMULATE))
         accumulates.push_back(requests[index]);
-      if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
+      if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
         requests[index] = MPI_REQUEST_NULL;
     }
     if (status != MPI_STATUSES_IGNORE) {
@@ -871,21 +1034,30 @@ int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
 {
   int count = 0;
+  int flag = 0;
+  int index = 0;
   MPI_Status stat;
   MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
 
+  index = waitany(incount, (MPI_Request*)requests, pstat);
+  if(index==MPI_UNDEFINED) return MPI_UNDEFINED;
+  if(status != MPI_STATUSES_IGNORE) {
+    status[count] = *pstat;
+  }
+  indices[count] = index;
+  count++;
   for (int i = 0; i < incount; i++) {
-    int index = waitany(incount, requests, pstat);
-    if(index!=MPI_UNDEFINED){
-      indices[count] = index;
-      count++;
-      if(status != MPI_STATUSES_IGNORE) {
-        status[index] = *pstat;
+    if((requests[i] != MPI_REQUEST_NULL)) {
+      test(&requests[i], pstat,&flag);
+      if (flag==1){
+        indices[count] = i;
+        if(status != MPI_STATUSES_IGNORE) {
+          status[count] = *pstat;
+        }
+        if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
+          requests[i]=MPI_REQUEST_NULL;
+        count++;
       }
-     if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
-       requests[index] = MPI_REQUEST_NULL;
-    }else{
-      return MPI_UNDEFINED;
     }
   }
   return count;
@@ -917,5 +1089,72 @@ void Request::free_f(int id)
   }
 }
 
+
+int Request::get_status(MPI_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);
+    if(*flag)
+      return MPI_SUCCESS;
+  }
+  if (req != MPI_REQUEST_NULL && 
+     (req->flags_ & MPI_REQ_GENERALIZED)
+     && !(req->flags_ & MPI_REQ_COMPLETE)) {
+     *flag=0;
+    return MPI_SUCCESS;
+  }
+
+  *flag=1;
+  if(req != MPI_REQUEST_NULL &&
+     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;
+    status->count = req->real_size_;
+  }
+  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){
+
+  *request = new Request();
+  (*request)->flags_ |= MPI_REQ_GENERALIZED;
+  (*request)->flags_ |= MPI_REQ_PERSISTENT;
+  (*request)->refcount_ = 1;
+  ((*request)->generalized_funcs)=xbt_new0(s_smpi_mpi_generalized_request_funcs_t ,1);
+  ((*request)->generalized_funcs)->query_fn=query_fn;
+  ((*request)->generalized_funcs)->free_fn=free_fn;
+  ((*request)->generalized_funcs)->cancel_fn=cancel_fn;
+  ((*request)->generalized_funcs)->extra_state=extra_state;
+  ((*request)->generalized_funcs)->cond = simgrid::s4u::ConditionVariable::create();
+  ((*request)->generalized_funcs)->mutex = simgrid::s4u::Mutex::create();
+  return MPI_SUCCESS;
+}
+
+int Request::grequest_complete( MPI_Request request){
+  if ((!(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex==NULL) 
+    return MPI_ERR_REQUEST;
+  request->generalized_funcs->mutex->lock();
+  request->flags_ |= MPI_REQ_COMPLETE; // in case wait would be called after complete
+  request->generalized_funcs->cond->notify_one();
+  request->generalized_funcs->mutex->unlock();
+  return MPI_SUCCESS;
+}
+
+void Request::set_nbc_requests(MPI_Request* reqs, int size){
+  nbc_requests_=reqs;
+  nbc_requests_size_=size;
+}
+
+int Request::get_nbc_requests_size(){
+  return nbc_requests_size_;
+}
+
+MPI_Request* Request::get_nbc_requests(){
+  return nbc_requests_;
+}
+
 }
 }