X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/9dca29757b5d8add4a2719c28f620686e7132556..4c753f8d4cabd4104f3f7109823f16be2ebdcce3:/src/smpi/bindings/smpi_pmpi_coll.cpp

diff --git a/src/smpi/bindings/smpi_pmpi_coll.cpp b/src/smpi/bindings/smpi_pmpi_coll.cpp
index 2a0a7458ab..f83a8d3c4b 100644
--- a/src/smpi/bindings/smpi_pmpi_coll.cpp
+++ b/src/smpi/bindings/smpi_pmpi_coll.cpp
@@ -1,4 +1,4 @@
-/* Copyright (c) 2007-2019. The SimGrid Team. All rights reserved.          */
+/* Copyright (c) 2007-2022. The SimGrid Team. All rights reserved.          */
 
 /* This program is free software; you can redistribute it and/or modify it
  * under the terms of the license (GNU LGPL) which comes with this package. */
@@ -11,19 +11,17 @@
 #include "smpi_op.hpp"
 #include "src/smpi/include/smpi_actor.hpp"
 
-XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(smpi_pmpi);
+#include <vector>
 
-#define CHECK_ARGS(test, errcode, ...)                                                                                 \
-  if (test) {                                                                                                          \
-    XBT_WARN(__VA_ARGS__);                                                                                             \
-    return (errcode);                                                                                                  \
-  }
+XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(smpi_pmpi);
 
-  static const void* smpi_get_in_place_buf(const void* inplacebuf, const void* otherbuf,std::unique_ptr<unsigned char[]>& tmp_sendbuf, int count, MPI_Datatype datatype){
+static const void* smpi_get_in_place_buf(const void* inplacebuf, const void* otherbuf,
+                                         std::vector<unsigned char>& tmp_sendbuf, int count, MPI_Datatype datatype)
+{
   if (inplacebuf == MPI_IN_PLACE) {
-      tmp_sendbuf.reset(new unsigned char[count * datatype->get_extent()]);
-      simgrid::smpi::Datatype::copy(otherbuf, count, datatype, tmp_sendbuf.get(), count, datatype);
-    return tmp_sendbuf.get();
+    tmp_sendbuf.resize(count * datatype->get_extent());
+    simgrid::smpi::Datatype::copy(otherbuf, count, datatype, tmp_sendbuf.data(), count, datatype);
+    return tmp_sendbuf.data();
   }else{
     return inplacebuf;
   }
@@ -37,14 +35,12 @@ int PMPI_Barrier(MPI_Comm comm)
 
 int PMPI_Ibarrier(MPI_Comm comm, MPI_Request *request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-
-  smpi_bench_end();
-  int rank = simgrid::s4u::this_actor::get_pid();
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Barrier" : "PMPI_Ibarrier",
+  CHECK_COMM(1)
+  CHECK_REQUEST(2)
+  CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Barrier" : "PMPI_Ibarrier")
+  const SmpiBenchGuard suspend_bench;
+  aid_t pid = simgrid::s4u::this_actor::get_pid();
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Barrier" : "PMPI_Ibarrier",
                      new simgrid::instr::NoOpTIData(request == MPI_REQUEST_IGNORED ? "barrier" : "ibarrier"));
   if (request == MPI_REQUEST_IGNORED) {
     simgrid::smpi::colls::barrier(comm);
@@ -53,8 +49,7 @@ int PMPI_Ibarrier(MPI_Comm comm, MPI_Request *request)
   } else
     simgrid::smpi::colls::ibarrier(comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return MPI_SUCCESS;
 }
 
@@ -63,28 +58,28 @@ int PMPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm c
   return PMPI_Ibcast(buf, count, datatype, root, comm, MPI_REQUEST_IGNORED);
 }
 
-int PMPI_Ibcast(void *buf, int count, MPI_Datatype datatype, 
-                   int root, MPI_Comm comm, MPI_Request* request)
+int PMPI_Ibcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm, MPI_Request* request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if (buf == nullptr && count > 0)
-    return MPI_ERR_BUFFER;
-  if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid())
-    return MPI_ERR_TYPE;
-  if (count < 0)
-    return MPI_ERR_COUNT;
-  if (root < 0 || root >= comm->size())
-    return MPI_ERR_ROOT;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-
-  smpi_bench_end();
-  int rank = simgrid::s4u::this_actor::get_pid();
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Bcast" : "PMPI_Ibcast",
+  SET_BUF1(buf)
+  CHECK_COMM(5)
+  CHECK_COUNT(2, count)
+  CHECK_TYPE(3, datatype)
+  CHECK_BUFFER(1, buf, count, datatype)
+  CHECK_ROOT(4)
+  CHECK_REQUEST(6)
+  CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Bcast" : "PMPI_Ibcast") + " with root " +
+                             std::to_string(root))
+
+  const SmpiBenchGuard suspend_bench;
+  aid_t pid = simgrid::s4u::this_actor::get_pid();
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Bcast" : "PMPI_Ibcast",
                      new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "bcast" : "ibcast", root, -1.0,
-                                                    datatype->is_replayable() ? count : count * datatype->size(), -1,
+                                                    count, 0,
                                                     simgrid::smpi::Datatype::encode(datatype), ""));
+  if (simgrid::config::get_value<bool>("smpi/barrier-collectives") &&
+      request == MPI_REQUEST_IGNORED) // No barrier in Ibcast
+    smpi_deployment_startup_barrier(smpi_process()->get_instance_id());
+
   if (comm->size() > 1) {
     if (request == MPI_REQUEST_IGNORED)
       simgrid::smpi::colls::bcast(buf, count, datatype, root, comm);
@@ -95,8 +90,7 @@ int PMPI_Ibcast(void *buf, int count, MPI_Datatype datatype,
       *request = MPI_REQUEST_NULL;
   }
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return MPI_SUCCESS;
 }
 
@@ -108,35 +102,53 @@ int PMPI_Gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,void *
 int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
                  MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if ((sendbuf == nullptr && sendcount > 0) || ((comm->rank() == root) && recvbuf == nullptr && recvcount > 0))
-    return MPI_ERR_BUFFER;
-  if (((sendbuf != MPI_IN_PLACE && sendcount > 0) && (sendtype == MPI_DATATYPE_NULL)) ||
-      ((comm->rank() == root) && (recvtype == MPI_DATATYPE_NULL)))
-    return MPI_ERR_TYPE;
-  if (((sendbuf != MPI_IN_PLACE) && (sendcount < 0)) || ((comm->rank() == root) && (recvcount < 0)))
-    return MPI_ERR_COUNT;
-  if (root < 0 || root >= comm->size())
-    return MPI_ERR_ROOT;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-
-  smpi_bench_end();
+  CHECK_COMM(8)
+  SET_BUF1(sendbuf)
+  int rank = comm->rank();
+  if(sendbuf != MPI_IN_PLACE){
+    CHECK_COUNT(2, sendcount)
+    CHECK_TYPE(3, sendtype)
+    CHECK_BUFFER(1,sendbuf, sendcount, sendtype)
+  }
+  if(rank == root){
+    SET_BUF2(recvbuf)
+    CHECK_NOT_IN_PLACE_ROOT(4, recvbuf)
+    CHECK_TYPE(6, recvtype)
+    CHECK_COUNT(5, recvcount)
+    CHECK_BUFFER(4, recvbuf, recvcount, recvtype)
+  } else {
+    CHECK_NOT_IN_PLACE_ROOT(1, sendbuf)
+  }
+  CHECK_ROOT(7)
+  CHECK_REQUEST(9)
+  CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Gather" : "PMPI_Igather") + " with root " +
+                             std::to_string(root))
+
   const void* real_sendbuf   = sendbuf;
   int real_sendcount         = sendcount;
   MPI_Datatype real_sendtype = sendtype;
-  if ((comm->rank() == root) && (sendbuf == MPI_IN_PLACE)) {
-    real_sendcount = 0;
-    real_sendtype  = recvtype;
+  if (rank == root){
+    if (sendbuf == MPI_IN_PLACE) {
+      real_sendcount = 0;
+      real_sendtype  = recvtype;
+    } else if(recvtype->size() * recvcount !=  sendtype->size() * sendcount){
+      XBT_WARN("MPI_(I)Gather : received size at root differs from sent size : %zu vs %zu", recvtype->size() * recvcount , sendtype->size() * sendcount);
+      return MPI_ERR_TRUNCATE;
+    }
   }
-  int rank = simgrid::s4u::this_actor::get_pid();
 
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Gather" : "PMPI_Igather",
+  const SmpiBenchGuard suspend_bench;
+
+  if (simgrid::config::get_value<bool>("smpi/barrier-collectives") &&
+      request == MPI_REQUEST_IGNORED) // no barrier in Igather
+    smpi_deployment_startup_barrier(smpi_process()->get_instance_id());
+
+  aid_t pid = simgrid::s4u::this_actor::get_pid();
+
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Gather" : "PMPI_Igather",
                      new simgrid::instr::CollTIData(
                          request == MPI_REQUEST_IGNORED ? "gather" : "igather", root, -1.0,
-                         real_sendtype->is_replayable() ? real_sendcount : real_sendcount * real_sendtype->size(),
-                         (comm->rank() != root || recvtype->is_replayable()) ? recvcount : recvcount * recvtype->size(),
+                         real_sendcount, recvcount,
                          simgrid::smpi::Datatype::encode(real_sendtype), simgrid::smpi::Datatype::encode(recvtype)));
   if (request == MPI_REQUEST_IGNORED)
     simgrid::smpi::colls::gather(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, root, comm);
@@ -144,8 +156,7 @@ int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void
     simgrid::smpi::colls::igather(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, root, comm,
                                   request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return MPI_SUCCESS;
 }
 
@@ -157,50 +168,62 @@ int PMPI_Gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void
 int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int* recvcounts, const int* displs,
                   MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if ((sendbuf == nullptr && sendcount > 0) || ((comm->rank() == root) && recvbuf == nullptr))
-    return MPI_ERR_BUFFER;
-  if (((sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) ||
-      ((comm->rank() == root) && (recvtype == MPI_DATATYPE_NULL)))
-    return MPI_ERR_TYPE;
-  if ((sendbuf != MPI_IN_PLACE) && (sendcount < 0))
-    return MPI_ERR_COUNT;
-  if ((comm->rank() == root) && (recvcounts == nullptr || displs == nullptr))
-    return MPI_ERR_ARG;
-  if (root < 0 || root >= comm->size())
-    return MPI_ERR_ROOT;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-
-  for (int i = 0; i < comm->size(); i++) {
-    if ((comm->rank() == root) && (recvcounts[i] < 0))
-      return MPI_ERR_COUNT;
+  CHECK_COMM(9)
+  SET_BUF1(sendbuf)
+  int rank = comm->rank();
+  if(sendbuf != MPI_IN_PLACE){
+    CHECK_TYPE(3, sendtype)
+    CHECK_COUNT(2, sendcount)
   }
+  CHECK_BUFFER(1, sendbuf, sendcount, sendtype)
+  if(rank == root){
+    SET_BUF2(recvbuf)
+    CHECK_NOT_IN_PLACE_ROOT(4, recvbuf)
+    CHECK_TYPE(6, recvtype)
+    CHECK_NULL(5, MPI_ERR_COUNT, recvcounts)
+    CHECK_NULL(6, MPI_ERR_ARG, displs)
+  } else {
+    CHECK_NOT_IN_PLACE_ROOT(1, sendbuf)
+  }
+  CHECK_ROOT(8)
+  CHECK_REQUEST(10)
+  CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Gatherv" : "PMPI_Igatherv") +
+                             " with root " + std::to_string(root))
+
+  if (rank == root){
+    for (int i = 0; i < comm->size(); i++) {
+      CHECK_COUNT(5, recvcounts[i])
+      CHECK_BUFFER(4,recvbuf,recvcounts[i], recvtype)
+    }
+  }
+
+  const SmpiBenchGuard suspend_bench;
+
+  if (simgrid::config::get_value<bool>("smpi/barrier-collectives") &&
+      request == MPI_REQUEST_IGNORED) // no barrier in Igatherv
+    smpi_deployment_startup_barrier(smpi_process()->get_instance_id());
 
-  smpi_bench_end();
   const void* real_sendbuf   = sendbuf;
   int real_sendcount         = sendcount;
   MPI_Datatype real_sendtype = sendtype;
-  if ((comm->rank() == root) && (sendbuf == MPI_IN_PLACE)) {
+  if ((rank == root) && (sendbuf == MPI_IN_PLACE)) {
     real_sendcount = 0;
     real_sendtype  = recvtype;
   }
 
-  int rank         = simgrid::s4u::this_actor::get_pid();
-  int dt_size_recv = recvtype->is_replayable() ? 1 : recvtype->size();
+  aid_t pid        = simgrid::s4u::this_actor::get_pid();
 
-  std::vector<int>* trace_recvcounts = new std::vector<int>;
-  if (comm->rank() == root) {
-    for (int i = 0; i < comm->size(); i++) // copy data to avoid bad free
-      trace_recvcounts->push_back(recvcounts[i] * dt_size_recv);
-  }
+  auto trace_recvcounts = std::make_shared<std::vector<int>>();
+  if (rank == root)
+    trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[comm->size()]);
+  else //this is not significant outside of root, put 0 as we don't know if recvcounts is initialized
+    trace_recvcounts->insert(trace_recvcounts->end(), comm->size(), 0);
 
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Gatherv" : "PMPI_Igatherv",
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Gatherv" : "PMPI_Igatherv",
                      new simgrid::instr::VarCollTIData(
                          request == MPI_REQUEST_IGNORED ? "gatherv" : "igatherv", root,
-                         real_sendtype->is_replayable() ? real_sendcount : real_sendcount * real_sendtype->size(),
-                         nullptr, dt_size_recv, trace_recvcounts, simgrid::smpi::Datatype::encode(real_sendtype),
+                         sendcount,
+                         nullptr, -1, trace_recvcounts, simgrid::smpi::Datatype::encode(real_sendtype),
                          simgrid::smpi::Datatype::encode(recvtype)));
   if (request == MPI_REQUEST_IGNORED)
     simgrid::smpi::colls::gatherv(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcounts, displs, recvtype,
@@ -209,8 +232,7 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi
     simgrid::smpi::colls::igatherv(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcounts, displs, recvtype,
                                    root, comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return MPI_SUCCESS;
 }
 
@@ -222,39 +244,48 @@ int PMPI_Allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
 int PMPI_Iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
                     MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request)
 {
-  CHECK_ARGS(comm == MPI_COMM_NULL, MPI_ERR_COMM, "Iallgather: the communicator cannot be MPI_COMM_NULL");
-  CHECK_ARGS(recvbuf == nullptr && recvcount > 0, MPI_ERR_BUFFER, "Iallgather: param 4 recvbuf cannot be NULL");
-  CHECK_ARGS(sendbuf == nullptr && sendcount > 0, MPI_ERR_BUFFER,
-             "Iallgather: param 1 sendbuf cannot be NULL when sendcound > 0");
-  CHECK_ARGS((sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL), MPI_ERR_TYPE,
-             "Iallgather: param 3 sendtype cannot be MPI_DATATYPE_NULL when sendbuff is not MPI_IN_PLACE");
-  CHECK_ARGS(recvtype == MPI_DATATYPE_NULL, MPI_ERR_TYPE, "Iallgather: param 6 recvtype cannot be MPI_DATATYPE_NULL");
-  CHECK_ARGS(recvcount < 0, MPI_ERR_COUNT, "Iallgather: param 5 recvcount cannot be negative");
-  CHECK_ARGS((sendbuf != MPI_IN_PLACE) && (sendcount < 0), MPI_ERR_COUNT,
-             "Iallgather: param 2 sendcount cannot be negative when sendbuf is not MPI_IN_PLACE");
-  CHECK_ARGS(request == nullptr, MPI_ERR_ARG, "Iallgather: param 8 request cannot be NULL");
-
-  smpi_bench_end();
+  CHECK_COMM(7)
+  SET_BUF1(sendbuf)
+  SET_BUF2(recvbuf)
+  int rank = comm->rank();
+  CHECK_NOT_IN_PLACE(4, recvbuf)
+  if(sendbuf != MPI_IN_PLACE){
+    CHECK_COUNT(2, sendcount)
+    CHECK_TYPE(3, sendtype)
+  }
+  CHECK_TYPE(6, recvtype)
+  CHECK_COUNT(5, recvcount)
+  CHECK_BUFFER(1, sendbuf, sendcount, sendtype)
+  CHECK_BUFFER(4, recvbuf, recvcount, recvtype)
+  CHECK_REQUEST(8)
+  CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Allgather" : "PMPI_Iallggather")
+
   if (sendbuf == MPI_IN_PLACE) {
     sendbuf   = static_cast<char*>(recvbuf) + recvtype->get_extent() * recvcount * comm->rank();
     sendcount = recvcount;
     sendtype  = recvtype;
   }
-  int rank = simgrid::s4u::this_actor::get_pid();
 
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Allgather" : "PMPI_Iallggather",
+  if(recvtype->size() * recvcount !=  sendtype->size() * sendcount){
+    XBT_WARN("MPI_(I)Allgather : received size from each process differs from sent size : %zu vs %zu", recvtype->size() * recvcount, sendtype->size() * sendcount);
+    return MPI_ERR_TRUNCATE;
+  }
+
+  const SmpiBenchGuard suspend_bench;
+
+  aid_t pid = simgrid::s4u::this_actor::get_pid();
+
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Allgather" : "PMPI_Iallggather",
                      new simgrid::instr::CollTIData(
                          request == MPI_REQUEST_IGNORED ? "allgather" : "iallgather", -1, -1.0,
-                         sendtype->is_replayable() ? sendcount : sendcount * sendtype->size(),
-                         recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(),
+                         sendcount, recvcount,
                          simgrid::smpi::Datatype::encode(sendtype), simgrid::smpi::Datatype::encode(recvtype)));
   if (request == MPI_REQUEST_IGNORED)
     simgrid::smpi::colls::allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
   else
     simgrid::smpi::colls::iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return MPI_SUCCESS;
 }
 
@@ -266,45 +297,44 @@ int PMPI_Allgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
 int PMPI_Iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int* recvcounts, const int* displs,
                      MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if (sendbuf == nullptr && sendcount > 0)
-    return MPI_ERR_BUFFER;
-  if (((sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || (recvtype == MPI_DATATYPE_NULL))
-    return MPI_ERR_TYPE;
-  if ((sendbuf != MPI_IN_PLACE) && (sendcount < 0))
-    return MPI_ERR_COUNT;
-  if (recvcounts == nullptr || displs == nullptr)
-    return MPI_ERR_ARG;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-
+  CHECK_COMM(8)
+  SET_BUF1(sendbuf)
+  SET_BUF2(recvbuf)
+  int rank = comm->rank();
+  if(sendbuf != MPI_IN_PLACE)
+    CHECK_TYPE(3, sendtype)
+  CHECK_TYPE(6, recvtype)
+  CHECK_NULL(5, MPI_ERR_COUNT, recvcounts)
+  CHECK_NULL(6, MPI_ERR_ARG, displs)
+  if(sendbuf != MPI_IN_PLACE){
+    CHECK_COUNT(2, sendcount)
+    CHECK_BUFFER(1, sendbuf, sendcount, sendtype)
+  }
+  CHECK_REQUEST(9)
+  CHECK_NOT_IN_PLACE(4, recvbuf)
   for (int i = 0; i < comm->size(); i++) {
-    if (recvcounts[i] < 0)
-      return MPI_ERR_COUNT;
-    else if (recvcounts[i] > 0 && recvbuf == nullptr)
-      return MPI_ERR_BUFFER;
+    CHECK_COUNT(5, recvcounts[i])
+    CHECK_BUFFER(4, recvbuf, recvcounts[i], recvtype)
   }
+  CHECK_COLLECTIVE(comm, MPI_REQUEST_IGNORED ? "PMPI_Allgatherv" : "PMPI_Iallgatherv")
+
+  const SmpiBenchGuard suspend_bench;
 
-  smpi_bench_end();
   if (sendbuf == MPI_IN_PLACE) {
     sendbuf   = static_cast<char*>(recvbuf) + recvtype->get_extent() * displs[comm->rank()];
     sendcount = recvcounts[comm->rank()];
     sendtype  = recvtype;
   }
-  int rank         = simgrid::s4u::this_actor::get_pid();
-  int dt_size_recv = recvtype->is_replayable() ? 1 : recvtype->size();
+  aid_t pid        = simgrid::s4u::this_actor::get_pid();
 
-  std::vector<int>* trace_recvcounts = new std::vector<int>;
-  for (int i = 0; i < comm->size(); i++) { // copy data to avoid bad free
-    trace_recvcounts->push_back(recvcounts[i] * dt_size_recv);
-  }
+  auto trace_recvcounts = std::make_shared<std::vector<int>>();
+  trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[comm->size()]);
 
   TRACE_smpi_comm_in(
-      rank, request == MPI_REQUEST_IGNORED ? "PMPI_Allgatherv" : "PMPI_Iallgatherv",
+      pid, request == MPI_REQUEST_IGNORED ? "PMPI_Allgatherv" : "PMPI_Iallgatherv",
       new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "allgatherv" : "iallgatherv", -1,
-                                        sendtype->is_replayable() ? sendcount : sendcount * sendtype->size(), nullptr,
-                                        dt_size_recv, trace_recvcounts, simgrid::smpi::Datatype::encode(sendtype),
+                                        sendcount, nullptr,
+                                        -1, trace_recvcounts, simgrid::smpi::Datatype::encode(sendtype),
                                         simgrid::smpi::Datatype::encode(recvtype)));
   if (request == MPI_REQUEST_IGNORED)
     simgrid::smpi::colls::allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm);
@@ -312,8 +342,7 @@ int PMPI_Iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype,
     simgrid::smpi::colls::iallgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm,
                                       request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return MPI_SUCCESS;
 }
 
@@ -325,41 +354,57 @@ int PMPI_Scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
 int PMPI_Iscatter(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
                   MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if (((comm->rank() == root) && (sendtype == MPI_DATATYPE_NULL || not sendtype->is_valid())) ||
-      ((recvbuf != MPI_IN_PLACE) && (recvtype == MPI_DATATYPE_NULL || not recvtype->is_valid())))
-    return MPI_ERR_TYPE;
-  if (((comm->rank() == root) && (sendcount < 0)) || ((recvbuf != MPI_IN_PLACE) && (recvcount < 0)))
-    return MPI_ERR_COUNT;
-  if ((sendbuf == recvbuf) || ((comm->rank() == root) && sendcount > 0 && (sendbuf == nullptr)) ||
-      (recvcount > 0 && recvbuf == nullptr))
-    return MPI_ERR_BUFFER;
-  if (root < 0 || root >= comm->size())
-    return MPI_ERR_ROOT;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-
-  smpi_bench_end();
+  CHECK_COMM(8)
+  SET_BUF2(recvbuf)
+  int rank = comm->rank();
+  if(rank == root){
+    SET_BUF1(sendbuf)
+    CHECK_NOT_IN_PLACE_ROOT(1, sendbuf)
+    CHECK_COUNT(2, sendcount)
+    CHECK_TYPE(3, sendtype)
+    CHECK_BUFFER(1, sendbuf, sendcount, sendtype)
+  } else {
+    CHECK_NOT_IN_PLACE_ROOT(4, recvbuf)
+  }
+  if(recvbuf != MPI_IN_PLACE){
+    CHECK_COUNT(5, recvcount)
+    CHECK_TYPE(6, recvtype)
+    CHECK_BUFFER(4, recvbuf, recvcount, recvtype)
+  }
+  CHECK_ROOT(8)
+  CHECK_REQUEST(9)
+  CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Scatter" : "PMPI_Iscatter") +
+                             " with root " + std::to_string(root))
+
   if (recvbuf == MPI_IN_PLACE) {
     recvtype  = sendtype;
     recvcount = sendcount;
   }
-  int rank = simgrid::s4u::this_actor::get_pid();
 
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Scatter" : "PMPI_Iscatter",
+  if((rank == root) && (recvtype->size() * recvcount !=  sendtype->size() * sendcount)){
+    XBT_WARN("MPI_(I)Scatter : sent size to each process differs from receive size");
+    return MPI_ERR_TRUNCATE;
+  }
+
+  const SmpiBenchGuard suspend_bench;
+
+  if (simgrid::config::get_value<bool>("smpi/barrier-collectives") &&
+      request == MPI_REQUEST_IGNORED) // no barrier in Iscatter
+    smpi_deployment_startup_barrier(smpi_process()->get_instance_id());
+
+  aid_t pid = simgrid::s4u::this_actor::get_pid();
+
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Scatter" : "PMPI_Iscatter",
                      new simgrid::instr::CollTIData(
                          request == MPI_REQUEST_IGNORED ? "scatter" : "iscatter", root, -1.0,
-                         (comm->rank() != root || sendtype->is_replayable()) ? sendcount : sendcount * sendtype->size(),
-                         recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(),
+                         sendcount, recvcount,
                          simgrid::smpi::Datatype::encode(sendtype), simgrid::smpi::Datatype::encode(recvtype)));
   if (request == MPI_REQUEST_IGNORED)
     simgrid::smpi::colls::scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm);
   else
     simgrid::smpi::colls::iscatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return MPI_SUCCESS;
 }
 
@@ -371,48 +416,55 @@ int PMPI_Scatterv(const void *sendbuf, const int *sendcounts, const int *displs,
 int PMPI_Iscatterv(const void* sendbuf, const int* sendcounts, const int* displs, MPI_Datatype sendtype, void* recvbuf, int recvcount,
                    MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request)
 {
-  CHECK_ARGS(comm == MPI_COMM_NULL, MPI_ERR_COMM, "Iscatterv: the communicator cannot be MPI_COMM_NULL");
-  CHECK_ARGS((comm->rank() == root) && (sendcounts == nullptr), MPI_ERR_ARG,
-             "Iscatterv: param 2 sendcounts cannot be NULL on the root rank");
-  CHECK_ARGS((comm->rank() == root) && (displs == nullptr), MPI_ERR_ARG,
-             "Iscatterv: param 3 displs cannot be NULL on the root rank");
-  CHECK_ARGS((comm->rank() == root) && (sendtype == MPI_DATATYPE_NULL), MPI_ERR_TYPE,
-             "Iscatterv: The sendtype cannot be NULL on the root rank");
-  CHECK_ARGS((recvbuf != MPI_IN_PLACE) && (recvtype == MPI_DATATYPE_NULL), MPI_ERR_TYPE,
-             "Iscatterv: the recvtype cannot be NULL when not receiving in place");
-  CHECK_ARGS(request == nullptr, MPI_ERR_ARG, "Iscatterv: param 10 request cannot be NULL");
-  CHECK_ARGS(recvbuf != MPI_IN_PLACE && recvcount < 0, MPI_ERR_COUNT,
-             "Iscatterv: When not receiving in place, the recvcount cannot be negative");
-  CHECK_ARGS(root < 0, MPI_ERR_ROOT, "Iscatterv: root cannot be negative");
-  CHECK_ARGS(root >= comm->size(), MPI_ERR_ROOT, "Iscatterv: root (=%d) is larger than communicator size (=%d)", root,
-             comm->size());
-
-  if (comm->rank() == root) {
+  SET_BUF2(recvbuf)
+  CHECK_COMM(9)
+  int rank = comm->rank();
+  if(recvbuf != MPI_IN_PLACE){
+    CHECK_COUNT(5, recvcount)
+    CHECK_TYPE(7, recvtype)
+    CHECK_BUFFER(4, recvbuf, recvcount, recvtype)
+  }
+  CHECK_ROOT(9)
+  CHECK_REQUEST(10)
+  if (rank == root) {
+    SET_BUF1(sendbuf)
+    CHECK_NOT_IN_PLACE_ROOT(1, sendbuf)
+    CHECK_NULL(2, MPI_ERR_COUNT, sendcounts)
+    CHECK_NULL(3, MPI_ERR_ARG, displs)
+    CHECK_TYPE(4, sendtype)
+    for (int i = 0; i < comm->size(); i++){
+      CHECK_COUNT(2, sendcounts[i])
+      CHECK_BUFFER(1, sendbuf, sendcounts[i], sendtype)
+    }
     if (recvbuf == MPI_IN_PLACE) {
       recvtype  = sendtype;
-      recvcount = sendcounts[comm->rank()];
+      recvcount = sendcounts[rank];
     }
-    for (int i = 0; i < comm->size(); i++)
-      CHECK_ARGS(sendcounts[i] < 0, MPI_ERR_COUNT, "Iscatterv: sendcounts[%d]=%d but this cannot be negative", i,
-                 sendcounts[i]);
+  } else {
+    CHECK_NOT_IN_PLACE_ROOT(4, recvbuf)
   }
+  CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Scatterv" : "PMPI_Iscatterv") +
+                             " with root " + std::to_string(root))
 
-  smpi_bench_end();
+  const SmpiBenchGuard suspend_bench;
 
-  int rank         = simgrid::s4u::this_actor::get_pid();
-  int dt_size_send = sendtype->is_replayable() ? 1 : sendtype->size();
+  if (simgrid::config::get_value<bool>("smpi/barrier-collectives") &&
+      request == MPI_REQUEST_IGNORED) // no barrier in Iscatterv
+    smpi_deployment_startup_barrier(smpi_process()->get_instance_id());
 
-  std::vector<int>* trace_sendcounts = new std::vector<int>;
-  if (comm->rank() == root) {
-    for (int i = 0; i < comm->size(); i++) { // copy data to avoid bad free
-      trace_sendcounts->push_back(sendcounts[i] * dt_size_send);
-    }
-  }
+  aid_t pid        = simgrid::s4u::this_actor::get_pid();
 
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Scatterv" : "PMPI_Iscatterv",
+  auto trace_sendcounts = std::make_shared<std::vector<int>>();
+  if (rank == root)
+    trace_sendcounts->insert(trace_sendcounts->end(), &sendcounts[0], &sendcounts[comm->size()]);
+  else //this is not significant outside of root, put 0 as we don't know if sendcounts is initialized
+    trace_sendcounts->insert(trace_sendcounts->end(), comm->size(), 0);
+
+
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Scatterv" : "PMPI_Iscatterv",
                      new simgrid::instr::VarCollTIData(
-                         request == MPI_REQUEST_IGNORED ? "scatterv" : "iscatterv", root, dt_size_send,
-                         trace_sendcounts, recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(),
+                         request == MPI_REQUEST_IGNORED ? "scatterv" : "iscatterv", root, -1,
+                         trace_sendcounts, recvcount,
                          nullptr, simgrid::smpi::Datatype::encode(sendtype),
                          simgrid::smpi::Datatype::encode(recvtype)));
   if (request == MPI_REQUEST_IGNORED)
@@ -421,8 +473,7 @@ int PMPI_Iscatterv(const void* sendbuf, const int* sendcounts, const int* displs
     simgrid::smpi::colls::iscatterv(sendbuf, sendcounts, displs, sendtype, recvbuf, recvcount, recvtype, root, comm,
                                     request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return MPI_SUCCESS;
 }
 
@@ -433,50 +484,56 @@ int PMPI_Reduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype data
 
 int PMPI_Ireduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm, MPI_Request* request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if ((sendbuf == nullptr && count > 0) || ((comm->rank() == root) && recvbuf == nullptr))
-    return MPI_ERR_BUFFER;
-  if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid())
-    return MPI_ERR_TYPE;
-  if (op == MPI_OP_NULL)
-    return MPI_ERR_OP;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-  if (root < 0 || root >= comm->size())
-    return MPI_ERR_ROOT;
-  if (count < 0)
-    return MPI_ERR_COUNT;
-
-  smpi_bench_end();
-  int rank = simgrid::s4u::this_actor::get_pid();
-
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce" : "PMPI_Ireduce",
+  CHECK_COMM(7)
+  SET_BUF1(sendbuf)
+  int rank = comm->rank();
+  CHECK_TYPE(4, datatype)
+  CHECK_COUNT(3, count)
+  CHECK_BUFFER(1, sendbuf, count, datatype)
+  if(rank == root){
+    SET_BUF2(recvbuf)
+    CHECK_NOT_IN_PLACE(2, recvbuf)
+    CHECK_BUFFER(5, recvbuf, count, datatype)
+  }
+  CHECK_OP(5, op, datatype)
+  CHECK_ROOT(7)
+  CHECK_REQUEST(8)
+  CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Reduce" : "PMPI_Ireduce") + " with op " +
+                             op->name() + " and root " + std::to_string(root))
+
+  const SmpiBenchGuard suspend_bench;
+
+  if (simgrid::config::get_value<bool>("smpi/barrier-collectives") &&
+      request == MPI_REQUEST_IGNORED) // no barrier in Ireduce
+    smpi_deployment_startup_barrier(smpi_process()->get_instance_id());
+
+  aid_t pid = simgrid::s4u::this_actor::get_pid();
+
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce" : "PMPI_Ireduce",
                      new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "reduce" : "ireduce", root, 0,
-                                                    datatype->is_replayable() ? count : count * datatype->size(), -1,
+                                                    count, 0,
                                                     simgrid::smpi::Datatype::encode(datatype), ""));
   if (request == MPI_REQUEST_IGNORED)
     simgrid::smpi::colls::reduce(sendbuf, recvbuf, count, datatype, op, root, comm);
   else
     simgrid::smpi::colls::ireduce(sendbuf, recvbuf, count, datatype, op, root, comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return MPI_SUCCESS;
 }
 
 int PMPI_Reduce_local(const void* inbuf, void* inoutbuf, int count, MPI_Datatype datatype, MPI_Op op)
 {
-  if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid())
-    return MPI_ERR_TYPE;
-  if (op == MPI_OP_NULL)
-    return MPI_ERR_OP;
-  if (count < 0)
-    return MPI_ERR_COUNT;
-
-  smpi_bench_end();
+  SET_BUF1(inbuf)
+  SET_BUF2(inoutbuf)
+  CHECK_TYPE(4, datatype)
+  CHECK_COUNT(3, count)
+  CHECK_BUFFER(1, inbuf, count, datatype)
+  CHECK_BUFFER(2, inoutbuf, count, datatype)
+  CHECK_OP(5, op, datatype)
+
+  const SmpiBenchGuard suspend_bench;
   op->apply(inbuf, inoutbuf, &count, datatype);
-  smpi_bench_begin();
   return MPI_SUCCESS;
 }
 
@@ -487,28 +544,30 @@ int PMPI_Allreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype d
 
 int PMPI_Iallreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if ((sendbuf == nullptr && count > 0) || (recvbuf == nullptr))
-    return MPI_ERR_BUFFER;
-  if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid())
-    return MPI_ERR_TYPE;
-  if (count < 0)
-    return MPI_ERR_COUNT;
-  if (op == MPI_OP_NULL)
-    return MPI_ERR_OP;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-
-  smpi_bench_end();
-  std::unique_ptr<unsigned char[]> tmp_sendbuf;
+  CHECK_COMM(6)
+  SET_BUF1(sendbuf)
+  SET_BUF2(recvbuf)
+  int rank = comm->rank();
+  CHECK_NOT_IN_PLACE(2, recvbuf)
+  CHECK_TYPE(4, datatype)
+  CHECK_OP(5, op, datatype)
+  CHECK_COUNT(3, count)
+  CHECK_BUFFER(1, sendbuf, count, datatype)
+  CHECK_BUFFER(2, recvbuf, count, datatype)
+  CHECK_REQUEST(7)
+  CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Alleduce" : "PMPI_Iallreduce") +
+                             " with op " + op->name())
+
+  const SmpiBenchGuard suspend_bench;
+
+  std::vector<unsigned char> tmp_sendbuf;
   const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype);
 
-  int rank = simgrid::s4u::this_actor::get_pid();
+  aid_t pid = simgrid::s4u::this_actor::get_pid();
 
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Allreduce" : "PMPI_Iallreduce",
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Allreduce" : "PMPI_Iallreduce",
                      new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "allreduce" : "iallreduce", -1, 0,
-                                                    datatype->is_replayable() ? count : count * datatype->size(), -1,
+                                                    count, 0,
                                                     simgrid::smpi::Datatype::encode(datatype), ""));
 
   if (request == MPI_REQUEST_IGNORED)
@@ -516,8 +575,7 @@ int PMPI_Iallreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype
   else
     simgrid::smpi::colls::iallreduce(real_sendbuf, recvbuf, count, datatype, op, comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return MPI_SUCCESS;
 }
 
@@ -528,28 +586,31 @@ int PMPI_Scan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype dataty
 
 int PMPI_Iscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request* request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid())
-    return MPI_ERR_TYPE;
-  if (op == MPI_OP_NULL)
-    return MPI_ERR_OP;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-  if (count < 0)
-    return MPI_ERR_COUNT;
-  if (sendbuf == nullptr || recvbuf == nullptr)
-    return MPI_ERR_BUFFER;
-
-  smpi_bench_end();
-  int rank         = simgrid::s4u::this_actor::get_pid();
-  std::unique_ptr<unsigned char[]> tmp_sendbuf;
+  CHECK_COMM(6)
+  SET_BUF1(sendbuf)
+  SET_BUF2(recvbuf)
+  CHECK_TYPE(4, datatype)
+  CHECK_COUNT(3, count)
+  CHECK_BUFFER(1,sendbuf,count, datatype)
+  CHECK_BUFFER(2,recvbuf,count, datatype)
+  CHECK_REQUEST(7)
+  CHECK_OP(5, op, datatype)
+  CHECK_COLLECTIVE(comm,
+                   std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Scan" : "PMPI_Iscan") + " with op " + op->name())
+
+  const SmpiBenchGuard suspend_bench;
+
+  if (simgrid::config::get_value<bool>("smpi/barrier-collectives") &&
+      request == MPI_REQUEST_IGNORED) // no barrier in Iscan
+    smpi_deployment_startup_barrier(smpi_process()->get_instance_id());
+
+  aid_t pid = simgrid::s4u::this_actor::get_pid();
+  std::vector<unsigned char> tmp_sendbuf;
   const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype);
 
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Scan" : "PMPI_Iscan",
-                     new simgrid::instr::Pt2PtTIData(request == MPI_REQUEST_IGNORED ? "scan" : "iscan", -1,
-                                                     datatype->is_replayable() ? count : count * datatype->size(),
-                                                     simgrid::smpi::Datatype::encode(datatype)));
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Scan" : "PMPI_Iscan",
+                     new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "scan" : "iscan", -1, 0.0,
+                                                    count, 0, simgrid::smpi::Datatype::encode(datatype), ""));
 
   int retval;
   if (request == MPI_REQUEST_IGNORED)
@@ -557,8 +618,7 @@ int PMPI_Iscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datat
   else
     retval = simgrid::smpi::colls::iscan(real_sendbuf, recvbuf, count, datatype, op, comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return retval;
 }
 
@@ -568,28 +628,31 @@ int PMPI_Exscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype data
 }
 
 int PMPI_Iexscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request* request){
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if (not datatype->is_valid())
-    return MPI_ERR_TYPE;
-  if (op == MPI_OP_NULL)
-    return MPI_ERR_OP;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-  if (count < 0)
-    return MPI_ERR_COUNT;
-  if (sendbuf == nullptr || recvbuf == nullptr)
-    return MPI_ERR_BUFFER;
-
-  smpi_bench_end();
-  int rank         = simgrid::s4u::this_actor::get_pid();
-  std::unique_ptr<unsigned char[]> tmp_sendbuf;
+  CHECK_COMM(6)
+  SET_BUF1(sendbuf)
+  SET_BUF2(recvbuf)
+  CHECK_TYPE(4, datatype)
+  CHECK_COUNT(3, count)
+  CHECK_BUFFER(1, sendbuf, count, datatype)
+  CHECK_BUFFER(2, recvbuf, count, datatype)
+  CHECK_REQUEST(7)
+  CHECK_OP(5, op, datatype)
+  CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Exscan" : "PMPI_Iexscan") + " with op " +
+                             op->name())
+
+  const SmpiBenchGuard suspend_bench;
+
+  if (simgrid::config::get_value<bool>("smpi/barrier-collectives") &&
+      request == MPI_REQUEST_IGNORED) // no barrier in Iexscan
+    smpi_deployment_startup_barrier(smpi_process()->get_instance_id());
+
+  aid_t pid = simgrid::s4u::this_actor::get_pid();
+  std::vector<unsigned char> tmp_sendbuf;
   const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype);
 
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Exscan" : "PMPI_Iexscan",
-                     new simgrid::instr::Pt2PtTIData(request == MPI_REQUEST_IGNORED ? "exscan" : "iexscan", -1,
-                                                     datatype->is_replayable() ? count : count * datatype->size(),
-                                                     simgrid::smpi::Datatype::encode(datatype)));
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Exscan" : "PMPI_Iexscan",
+                     new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "exscan" : "iexscan", -1, 0.0,
+                                                    count, 0, simgrid::smpi::Datatype::encode(datatype), ""));
 
   int retval;
   if (request == MPI_REQUEST_IGNORED)
@@ -597,8 +660,7 @@ int PMPI_Iexscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype dat
   else
     retval = simgrid::smpi::colls::iexscan(real_sendbuf, recvbuf, count, datatype, op, comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return retval;
 }
 
@@ -609,49 +671,48 @@ int PMPI_Reduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcount
 
 int PMPI_Ireduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcounts, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if ((sendbuf == nullptr) || (recvbuf == nullptr))
-    return MPI_ERR_BUFFER;
-  if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid())
-    return MPI_ERR_TYPE;
-  if (op == MPI_OP_NULL)
-    return MPI_ERR_OP;
-  if (recvcounts == nullptr)
-    return MPI_ERR_ARG;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-
-  for (int i = 0; i < comm->size(); i++) { // copy data to avoid bad free
-    if (recvcounts[i] < 0)
-      return MPI_ERR_COUNT;
+  CHECK_COMM(6)
+  SET_BUF1(sendbuf)
+  SET_BUF2(recvbuf)
+  int rank = comm->rank();
+  CHECK_NOT_IN_PLACE(2, recvbuf)
+  CHECK_TYPE(4, datatype)
+  CHECK_NULL(3, MPI_ERR_COUNT, recvcounts)
+  CHECK_REQUEST(7)
+  CHECK_OP(5, op, datatype)
+  for (int i = 0; i < comm->size(); i++) {
+    CHECK_COUNT(3, recvcounts[i])
+    CHECK_BUFFER(1, sendbuf, recvcounts[i], datatype)
+    CHECK_BUFFER(2, recvbuf, recvcounts[i], datatype)
   }
+  CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter" : "PMPI_Ireduce_scatter") +
+                             " with op " + op->name())
+
+  const SmpiBenchGuard suspend_bench;
+
+  aid_t pid                          = simgrid::s4u::this_actor::get_pid();
+  auto trace_recvcounts              = std::make_shared<std::vector<int>>();
+  trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[comm->size()]);
 
-  smpi_bench_end();
-  int rank                           = simgrid::s4u::this_actor::get_pid();
-  std::vector<int>* trace_recvcounts = new std::vector<int>;
-  int dt_send_size                   = datatype->is_replayable() ? 1 : datatype->size();
   int totalcount                     = 0;
 
   for (int i = 0; i < comm->size(); i++) { // copy data to avoid bad free
-    trace_recvcounts->push_back(recvcounts[i] * dt_send_size);
     totalcount += recvcounts[i];
   }
-  std::unique_ptr<unsigned char[]> tmp_sendbuf;
+  std::vector<unsigned char> tmp_sendbuf;
   const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, totalcount, datatype);
 
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter" : "PMPI_Ireduce_scatter",
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter" : "PMPI_Ireduce_scatter",
                      new simgrid::instr::VarCollTIData(
-                         request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, dt_send_size, nullptr,
-                         -1, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), ""));
+                         request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, -1, nullptr,
+                         -1 , trace_recvcounts, std::to_string(0), simgrid::smpi::Datatype::encode(datatype)));
 
   if (request == MPI_REQUEST_IGNORED)
     simgrid::smpi::colls::reduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm);
   else
     simgrid::smpi::colls::ireduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return MPI_SUCCESS;
 }
 
@@ -664,42 +725,43 @@ int PMPI_Reduce_scatter_block(const void *sendbuf, void *recvbuf, int recvcount,
 int PMPI_Ireduce_scatter_block(const void* sendbuf, void* recvbuf, int recvcount, MPI_Datatype datatype, MPI_Op op,
                                MPI_Comm comm, MPI_Request* request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if (not datatype->is_valid())
-    return MPI_ERR_TYPE;
-  if (op == MPI_OP_NULL)
-    return MPI_ERR_OP;
-  if (recvcount < 0)
-    return MPI_ERR_ARG;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-
-  smpi_bench_end();
+  CHECK_COMM(6)
+  SET_BUF1(sendbuf)
+  SET_BUF2(recvbuf)
+  CHECK_TYPE(4, datatype)
+  CHECK_COUNT(3, recvcount)
+  CHECK_BUFFER(1, sendbuf, recvcount, datatype)
+  CHECK_BUFFER(2, recvbuf, recvcount, datatype)
+  CHECK_REQUEST(7)
+  CHECK_OP(5, op, datatype)
+  CHECK_COLLECTIVE(
+      comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter_block" : "PMPI_Ireduce_scatter_block") +
+                " with op " + op->name())
+
+  const SmpiBenchGuard suspend_bench;
+
   int count = comm->size();
 
-  int rank                           = simgrid::s4u::this_actor::get_pid();
-  int dt_send_size                   = datatype->is_replayable() ? 1 : datatype->size();
-  std::vector<int>* trace_recvcounts = new std::vector<int>(recvcount * dt_send_size); // copy data to avoid bad free
-  std::unique_ptr<unsigned char[]> tmp_sendbuf;
+  aid_t pid                          = simgrid::s4u::this_actor::get_pid();
+  auto trace_recvcounts = std::make_shared<std::vector<int>>(recvcount);
+
+  std::vector<unsigned char> tmp_sendbuf;
   const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, recvcount * count, datatype);
 
   TRACE_smpi_comm_in(
-      rank, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter_block" : "PMPI_Ireduce_scatter_block",
-      new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, 0,
+      pid, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter_block" : "PMPI_Ireduce_scatter_block",
+      new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, -1,
                                         nullptr, -1, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), ""));
 
-  int* recvcounts = new int[count];
+  std::vector<int> recvcounts(count);
   for (int i      = 0; i < count; i++)
     recvcounts[i] = recvcount;
   if (request == MPI_REQUEST_IGNORED)
-    simgrid::smpi::colls::reduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm);
+    simgrid::smpi::colls::reduce_scatter(real_sendbuf, recvbuf, recvcounts.data(), datatype, op, comm);
   else
-    simgrid::smpi::colls::ireduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm, request);
-  delete[] recvcounts;
+    simgrid::smpi::colls::ireduce_scatter(real_sendbuf, recvbuf, recvcounts.data(), datatype, op, comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return MPI_SUCCESS;
 }
 
@@ -711,23 +773,26 @@ int PMPI_Alltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi
 int PMPI_Ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
                    MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if ((sendbuf == nullptr && sendcount > 0) || (recvbuf == nullptr && recvcount > 0))
-    return MPI_ERR_BUFFER;
-  if ((sendbuf != MPI_IN_PLACE && sendtype == MPI_DATATYPE_NULL) || recvtype == MPI_DATATYPE_NULL)
-    return MPI_ERR_TYPE;
-  if ((sendbuf != MPI_IN_PLACE && sendcount < 0) || recvcount < 0)
-    return MPI_ERR_COUNT;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-
-  smpi_bench_end();
-  int rank                 = simgrid::s4u::this_actor::get_pid();
+  CHECK_COMM(7)
+  SET_BUF1(sendbuf)
+  SET_BUF2(recvbuf)
+  if(sendbuf != MPI_IN_PLACE){
+    CHECK_TYPE(3, sendtype)
+    CHECK_COUNT(2, sendcount)
+    CHECK_BUFFER(1, sendbuf, sendcount, sendtype)
+  }
+  CHECK_TYPE(6, recvtype)
+  CHECK_COUNT(5, recvcount)
+  CHECK_COUNT(5, recvcount)
+  CHECK_BUFFER(4, recvbuf, recvcount, recvtype)
+  CHECK_REQUEST(8)
+  CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoall" : "PMPI_Ialltoall")
+
+  aid_t pid                  = simgrid::s4u::this_actor::get_pid();
   int real_sendcount         = sendcount;
   MPI_Datatype real_sendtype = sendtype;
-  
-  std::unique_ptr<unsigned char[]> tmp_sendbuf;
+
+  std::vector<unsigned char> tmp_sendbuf;
   const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, recvcount * comm->size(), recvtype);
 
   if (sendbuf == MPI_IN_PLACE) {
@@ -735,11 +800,17 @@ int PMPI_Ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, vo
     real_sendtype  = recvtype;
   }
 
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoall" : "PMPI_Ialltoall",
+  if(recvtype->size() * recvcount !=  real_sendtype->size() * real_sendcount){
+    XBT_WARN("MPI_(I)Alltoall : receive size from each process differs from sent size : %zu vs %zu", recvtype->size() * recvcount,  real_sendtype->size() * real_sendcount);
+    return MPI_ERR_TRUNCATE;
+  }
+
+  const SmpiBenchGuard suspend_bench;
+
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoall" : "PMPI_Ialltoall",
                      new simgrid::instr::CollTIData(
                          request == MPI_REQUEST_IGNORED ? "alltoall" : "ialltoall", -1, -1.0,
-                         real_sendtype->is_replayable() ? real_sendcount : real_sendcount * real_sendtype->size(),
-                         recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(),
+                         real_sendcount, recvcount,
                          simgrid::smpi::Datatype::encode(real_sendtype), simgrid::smpi::Datatype::encode(recvtype)));
   int retval;
   if (request == MPI_REQUEST_IGNORED)
@@ -749,79 +820,87 @@ int PMPI_Ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, vo
     retval = simgrid::smpi::colls::ialltoall(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype,
                                              comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return retval;
 }
 
-int PMPI_Alltoallv(const void* sendbuf, const int* sendcounts, const int* senddisps, MPI_Datatype sendtype, void* recvbuf,
-                   const int* recvcounts, const int* recvdisps, MPI_Datatype recvtype, MPI_Comm comm)
+int PMPI_Alltoallv(const void* sendbuf, const int* sendcounts, const int* senddispls, MPI_Datatype sendtype, void* recvbuf,
+                   const int* recvcounts, const int* recvdispls, MPI_Datatype recvtype, MPI_Comm comm)
 {
-  return PMPI_Ialltoallv(sendbuf, sendcounts, senddisps, sendtype, recvbuf, recvcounts, recvdisps, recvtype, comm, MPI_REQUEST_IGNORED);
+  return PMPI_Ialltoallv(sendbuf, sendcounts, senddispls, sendtype, recvbuf, recvcounts, recvdispls, recvtype, comm, MPI_REQUEST_IGNORED);
 }
 
-int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* senddisps, MPI_Datatype sendtype, void* recvbuf,
-                    const int* recvcounts, const int* recvdisps, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request)
+int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* senddispls, MPI_Datatype sendtype, void* recvbuf,
+                    const int* recvcounts, const int* recvdispls, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if (sendbuf == nullptr || recvbuf == nullptr)
-    return MPI_ERR_BUFFER;
-  if ((sendbuf != MPI_IN_PLACE && sendtype == MPI_DATATYPE_NULL) || recvtype == MPI_DATATYPE_NULL)
-    return MPI_ERR_TYPE;
-  if ((sendbuf != MPI_IN_PLACE && (sendcounts == nullptr || senddisps == nullptr)) || recvcounts == nullptr ||
-      recvdisps == nullptr)
-    return MPI_ERR_ARG;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-
-  int rank = simgrid::s4u::this_actor::get_pid();
+  CHECK_COMM(9)
+  SET_BUF1(sendbuf)
+  SET_BUF2(recvbuf)
+  if(sendbuf != MPI_IN_PLACE){
+    CHECK_NULL(2, MPI_ERR_COUNT, sendcounts)
+    CHECK_NULL(3, MPI_ERR_ARG, senddispls)
+    CHECK_TYPE(4, sendtype)
+  }
+  CHECK_TYPE(8, recvtype)
+  CHECK_NULL(6, MPI_ERR_COUNT, recvcounts)
+  CHECK_NULL(7, MPI_ERR_ARG, recvdispls)
+  CHECK_REQUEST(10)
+  CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallv" : "PMPI_Ialltoallv")
+
+  aid_t pid = simgrid::s4u::this_actor::get_pid();
   int size = comm->size();
   for (int i = 0; i < size; i++) {
-    if (recvcounts[i] < 0 || (sendbuf != MPI_IN_PLACE && sendcounts[i] < 0))
-      return MPI_ERR_COUNT;
+    if(sendbuf != MPI_IN_PLACE){
+      CHECK_BUFFER(1, sendbuf, sendcounts[i], sendtype)
+      CHECK_COUNT(2, sendcounts[i])
+    }
+    CHECK_BUFFER(5, recvbuf, recvcounts[i], recvtype)
+    CHECK_COUNT(6, recvcounts[i])
   }
 
-  smpi_bench_end();
+  const SmpiBenchGuard suspend_bench;
+
   int send_size                      = 0;
   int recv_size                      = 0;
-  std::vector<int>* trace_sendcounts = new std::vector<int>;
-  std::vector<int>* trace_recvcounts = new std::vector<int>;
+  auto trace_sendcounts              = std::make_shared<std::vector<int>>();
+  auto trace_recvcounts              = std::make_shared<std::vector<int>>();
+  trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[size]);
+
   int dt_size_recv                   = recvtype->size();
 
   const int* real_sendcounts = sendcounts;
-  const int* real_senddisps  = senddisps;
+  const int* real_senddispls  = senddispls;
   MPI_Datatype real_sendtype = sendtype;
   int maxsize              = 0;
-  for (int i = 0; i < size; i++) { // copy data to avoid bad free
-    recv_size += recvcounts[i] * dt_size_recv;
-    trace_recvcounts->push_back(recvcounts[i] * dt_size_recv);
-    if (((recvdisps[i] + recvcounts[i]) * dt_size_recv) > maxsize)
-      maxsize = (recvdisps[i] + recvcounts[i]) * dt_size_recv;
-  }
+  std::vector<unsigned char> tmp_sendbuf;
+  std::vector<int> tmp_sendcounts;
+  std::vector<int> tmp_senddispls;
+  const void* real_sendbuf;
 
-  std::unique_ptr<unsigned char[]> tmp_sendbuf;
-  std::unique_ptr<int[]> tmp_sendcounts;
-  std::unique_ptr<int[]> tmp_senddisps;
-  const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, maxsize, MPI_CHAR);
   if (sendbuf == MPI_IN_PLACE) {
-    tmp_sendcounts.reset(new int[size]);
-    std::copy(recvcounts, recvcounts + size, tmp_sendcounts.get());
-    real_sendcounts = tmp_sendcounts.get();
-    tmp_senddisps.reset(new int[size]);
-    std::copy(recvdisps, recvdisps + size, tmp_senddisps.get());
-    real_senddisps = tmp_senddisps.get();
+    tmp_sendcounts.assign(recvcounts, recvcounts + size);
+    real_sendcounts = tmp_sendcounts.data();
+    tmp_senddispls.assign(recvdispls, recvdispls + size);
+    real_senddispls = tmp_senddispls.data();
     real_sendtype  = recvtype;
   }
 
-  int dt_size_send = real_sendtype->size();
-
   for (int i = 0; i < size; i++) { // copy data to avoid bad free
-    send_size += real_sendcounts[i] * dt_size_send;
-    trace_sendcounts->push_back(real_sendcounts[i] * dt_size_send);
+    send_size += real_sendcounts[i] ;
+    recv_size += recvcounts[i];
+    if (((recvdispls[i] + recvcounts[i]) * dt_size_recv) > maxsize)
+      maxsize = (recvdispls[i] + recvcounts[i]) * dt_size_recv;
+  }
+  real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, maxsize, MPI_CHAR);
+
+  if(recvtype->size() * recvcounts[comm->rank()] !=  real_sendtype->size() * real_sendcounts[comm->rank()]){
+    XBT_WARN("MPI_(I)Alltoallv : receive size from me differs from sent size to me : %zu vs %zu", recvtype->size() * recvcounts[comm->rank()], real_sendtype->size() * real_sendcounts[comm->rank()]);
+    return MPI_ERR_TRUNCATE;
   }
 
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallv" : "PMPI_Ialltoallv",
+  trace_sendcounts->insert(trace_sendcounts->end(), &real_sendcounts[0], &real_sendcounts[size]);
+
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallv" : "PMPI_Ialltoallv",
                      new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "alltoallv" : "ialltoallv", -1,
                                                        send_size, trace_sendcounts, recv_size, trace_recvcounts,
                                                        simgrid::smpi::Datatype::encode(real_sendtype),
@@ -829,89 +908,98 @@ int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* sendd
 
   int retval;
   if (request == MPI_REQUEST_IGNORED)
-    retval = simgrid::smpi::colls::alltoallv(real_sendbuf, real_sendcounts, real_senddisps, real_sendtype, recvbuf,
-                                             recvcounts, recvdisps, recvtype, comm);
+    retval = simgrid::smpi::colls::alltoallv(real_sendbuf, real_sendcounts, real_senddispls, real_sendtype, recvbuf,
+                                             recvcounts, recvdispls, recvtype, comm);
   else
-    retval = simgrid::smpi::colls::ialltoallv(real_sendbuf, real_sendcounts, real_senddisps, real_sendtype, recvbuf,
-                                              recvcounts, recvdisps, recvtype, comm, request);
+    retval = simgrid::smpi::colls::ialltoallv(real_sendbuf, real_sendcounts, real_senddispls, real_sendtype, recvbuf,
+                                              recvcounts, recvdispls, recvtype, comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return retval;
 }
 
-int PMPI_Alltoallw(const void* sendbuf, const int* sendcounts, const int* senddisps, const MPI_Datatype* sendtypes, void* recvbuf,
-                   const int* recvcounts, const int* recvdisps, const MPI_Datatype* recvtypes, MPI_Comm comm)
+int PMPI_Alltoallw(const void* sendbuf, const int* sendcounts, const int* senddispls, const MPI_Datatype* sendtypes, void* recvbuf,
+                   const int* recvcounts, const int* recvdispls, const MPI_Datatype* recvtypes, MPI_Comm comm)
 {
-  return PMPI_Ialltoallw(sendbuf, sendcounts, senddisps, sendtypes, recvbuf, recvcounts, recvdisps, recvtypes, comm, MPI_REQUEST_IGNORED);
+  return PMPI_Ialltoallw(sendbuf, sendcounts, senddispls, sendtypes, recvbuf, recvcounts, recvdispls, recvtypes, comm, MPI_REQUEST_IGNORED);
 }
 
-int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* senddisps, const MPI_Datatype* sendtypes, void* recvbuf,
-                    const int* recvcounts, const int* recvdisps, const MPI_Datatype* recvtypes, MPI_Comm comm, MPI_Request* request)
+int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* senddispls, const MPI_Datatype* sendtypes, void* recvbuf,
+                    const int* recvcounts, const int* recvdispls, const MPI_Datatype* recvtypes, MPI_Comm comm, MPI_Request* request)
 {
-  if (comm == MPI_COMM_NULL)
-    return MPI_ERR_COMM;
-  if (sendbuf == nullptr || recvbuf == nullptr)
-    return MPI_ERR_BUFFER;
-  if ((sendbuf != MPI_IN_PLACE && sendtypes == nullptr) || recvtypes == nullptr)
-    return MPI_ERR_TYPE;
-  if ((sendbuf != MPI_IN_PLACE && (sendcounts == nullptr || senddisps == nullptr)) || recvcounts == nullptr ||
-      recvdisps == nullptr)
-    return MPI_ERR_ARG;
-  if (request == nullptr)
-    return MPI_ERR_ARG;
-
-  smpi_bench_end();
-  int rank = simgrid::s4u::this_actor::get_pid();
+  CHECK_COMM(9)
+  SET_BUF1(sendbuf)
+  SET_BUF2(recvbuf)
+  if(sendbuf != MPI_IN_PLACE){
+    CHECK_NULL(2, MPI_ERR_COUNT, sendcounts)
+    CHECK_NULL(3, MPI_ERR_ARG, senddispls)
+    CHECK_NULL(4, MPI_ERR_TYPE, sendtypes)
+  }
+  CHECK_NULL(6, MPI_ERR_COUNT, recvcounts)
+  CHECK_NULL(7, MPI_ERR_ARG, recvdispls)
+  CHECK_NULL(8, MPI_ERR_TYPE, recvtypes)
+  CHECK_REQUEST(10)
+  aid_t pid = simgrid::s4u::this_actor::get_pid();
   int size = comm->size();
   for (int i = 0; i < size; i++) {
-    if (recvcounts[i] < 0 || (sendbuf != MPI_IN_PLACE && sendcounts[i] < 0))
-      return MPI_ERR_COUNT;
+    if(sendbuf != MPI_IN_PLACE){
+      CHECK_COUNT(2, sendcounts[i])
+      CHECK_TYPE(4, sendtypes[i])
+      CHECK_BUFFER(1, sendbuf, sendcounts[i], sendtypes[i])
+    }
+    CHECK_COUNT(6, recvcounts[i])
+    CHECK_TYPE(8, recvtypes[i])
+    CHECK_BUFFER(5, recvbuf, recvcounts[i], recvtypes[i])
   }
+  CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallw" : "PMPI_Ialltoallw")
+
+  const SmpiBenchGuard suspend_bench;
+
   int send_size                      = 0;
   int recv_size                      = 0;
-  std::vector<int>* trace_sendcounts = new std::vector<int>;
-  std::vector<int>* trace_recvcounts = new std::vector<int>;
+  auto trace_sendcounts              = std::make_shared<std::vector<int>>();
+  auto trace_recvcounts              = std::make_shared<std::vector<int>>();
+  trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[size]);
 
   const int* real_sendcounts         = sendcounts;
-  const int* real_senddisps          = senddisps;
+  const int* real_senddispls          = senddispls;
   const MPI_Datatype* real_sendtypes = sendtypes;
+
   unsigned long maxsize      = 0;
   for (int i = 0; i < size; i++) { // copy data to avoid bad free
-    if (recvtypes[i] == MPI_DATATYPE_NULL) {
-      delete trace_recvcounts;
-      delete trace_sendcounts;
+    if (recvtypes[i] == MPI_DATATYPE_NULL)
       return MPI_ERR_TYPE;
-    }
     recv_size += recvcounts[i] * recvtypes[i]->size();
-    trace_recvcounts->push_back(recvcounts[i] * recvtypes[i]->size());
-    if ((recvdisps[i] + (recvcounts[i] * recvtypes[i]->size())) > maxsize)
-      maxsize = recvdisps[i] + (recvcounts[i] * recvtypes[i]->size());
+    if ((recvdispls[i] + (recvcounts[i] * recvtypes[i]->size())) > maxsize)
+      maxsize = recvdispls[i] + (recvcounts[i] * recvtypes[i]->size());
   }
 
-  std::unique_ptr<unsigned char[]> tmp_sendbuf;
-  std::unique_ptr<int[]> tmp_sendcounts;
-  std::unique_ptr<int[]> tmp_senddisps;
-  std::unique_ptr<MPI_Datatype[]> tmp_sendtypes;
+  std::vector<unsigned char> tmp_sendbuf;
+  std::vector<int> tmp_sendcounts;
+  std::vector<int> tmp_senddispls;
+  std::vector<MPI_Datatype> tmp_sendtypes;
   const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, maxsize, MPI_CHAR);
   if (sendbuf == MPI_IN_PLACE) {
-    tmp_sendcounts.reset(new int[size]);
-    std::copy(recvcounts, recvcounts + size, tmp_sendcounts.get());
-    real_sendcounts = tmp_sendcounts.get();
-    tmp_senddisps.reset(new int[size]);
-    std::copy(recvdisps, recvdisps + size, tmp_senddisps.get());
-    real_senddisps = tmp_senddisps.get();
-    tmp_sendtypes.reset(new MPI_Datatype[size]);
-    std::copy(recvtypes, recvtypes + size, tmp_sendtypes.get());
-    real_sendtypes = tmp_sendtypes.get();
+    tmp_sendcounts.assign(recvcounts, recvcounts + size);
+    real_sendcounts = tmp_sendcounts.data();
+    tmp_senddispls.assign(recvdispls, recvdispls + size);
+    real_senddispls = tmp_senddispls.data();
+    tmp_sendtypes.assign(recvtypes, recvtypes + size);
+    real_sendtypes = tmp_sendtypes.data();
+  }
+
+
+  if(recvtypes[comm->rank()]->size() * recvcounts[comm->rank()] !=  real_sendtypes[comm->rank()]->size() * real_sendcounts[comm->rank()]){
+    XBT_WARN("MPI_(I)Alltoallw : receive size from me differs from sent size to me : %zu vs %zu", recvtypes[comm->rank()]->size() * recvcounts[comm->rank()],  real_sendtypes[comm->rank()]->size() * real_sendcounts[comm->rank()]);
+    return MPI_ERR_TRUNCATE;
   }
 
+  trace_sendcounts->insert(trace_sendcounts->end(), &real_sendcounts[0], &real_sendcounts[size]);
   for (int i = 0; i < size; i++) { // copy data to avoid bad free
     send_size += real_sendcounts[i] * real_sendtypes[i]->size();
-    trace_sendcounts->push_back(real_sendcounts[i] * real_sendtypes[i]->size());
   }
 
-  TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallw" : "PMPI_Ialltoallw",
+  TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallw" : "PMPI_Ialltoallw",
                      new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "alltoallv" : "ialltoallv", -1,
                                                        send_size, trace_sendcounts, recv_size, trace_recvcounts,
                                                        simgrid::smpi::Datatype::encode(real_sendtypes[0]),
@@ -919,13 +1007,12 @@ int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* sendd
 
   int retval;
   if (request == MPI_REQUEST_IGNORED)
-    retval = simgrid::smpi::colls::alltoallw(real_sendbuf, real_sendcounts, real_senddisps, real_sendtypes, recvbuf,
-                                             recvcounts, recvdisps, recvtypes, comm);
+    retval = simgrid::smpi::colls::alltoallw(real_sendbuf, real_sendcounts, real_senddispls, real_sendtypes, recvbuf,
+                                             recvcounts, recvdispls, recvtypes, comm);
   else
-    retval = simgrid::smpi::colls::ialltoallw(real_sendbuf, real_sendcounts, real_senddisps, real_sendtypes, recvbuf,
-                                              recvcounts, recvdisps, recvtypes, comm, request);
+    retval = simgrid::smpi::colls::ialltoallw(real_sendbuf, real_sendcounts, real_senddispls, real_sendtypes, recvbuf,
+                                              recvcounts, recvdispls, recvtypes, comm, request);
 
-  TRACE_smpi_comm_out(rank);
-  smpi_bench_begin();
+  TRACE_smpi_comm_out(pid);
   return retval;
 }