X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/f7d3e337e64b54125c8a990a99583383a6bf9780..8df941248c3c358bcfe0be644fc11b8a436ad5c5:/src/smpi/smpi_pmpi.c

diff --git a/src/smpi/smpi_pmpi.c b/src/smpi/smpi_pmpi.c
index 0d93e9c38a..5b51a9bf89 100644
--- a/src/smpi/smpi_pmpi.c
+++ b/src/smpi/smpi_pmpi.c
@@ -1,8 +1,8 @@
-/* Copyright (c) 2007-2013. The SimGrid Team.
+/* Copyright (c) 2007-2014. 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. */
+ * under the terms of the license (GNU LGPL) which comes with this package. */
 
 #include "private.h"
 #include "smpi_mpi_dt_private.h"
@@ -31,8 +31,11 @@ int PMPI_Init(int *argc, char ***argv)
 #ifdef HAVE_TRACING
   int rank = smpi_process_index();
   TRACE_smpi_init(rank);
-
   TRACE_smpi_computing_init(rank);
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_INIT;
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, extra);
+  TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
 #endif
   smpi_bench_begin();
   return MPI_SUCCESS;
@@ -40,11 +43,16 @@ int PMPI_Init(int *argc, char ***argv)
 
 int PMPI_Finalize(void)
 {
-  smpi_process_finalize();
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = smpi_process_index();
-  TRACE_smpi_computing_out(rank);
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_FINALIZE;
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, extra);
+#endif
+  smpi_process_finalize();
+#ifdef HAVE_TRACING
+  TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
   TRACE_smpi_finalize(smpi_process_index());
 #endif
   smpi_process_destroy();
@@ -85,14 +93,12 @@ int PMPI_Query_thread(int *provided)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (provided == NULL) {
     retval = MPI_ERR_ARG;
   } else {
     *provided = MPI_THREAD_MULTIPLE;
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -100,14 +106,12 @@ int PMPI_Is_thread_main(int *flag)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (flag == NULL) {
     retval = MPI_ERR_ARG;
   } else {
     *flag = smpi_process_index() == 0;
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -115,10 +119,6 @@ int PMPI_Abort(MPI_Comm comm, int errorcode)
 {
   smpi_bench_end();
   smpi_process_destroy();
-#ifdef HAVE_TRACING
-  int rank = smpi_process_index();
-  TRACE_smpi_computing_out(rank);
-#endif
   // FIXME: should kill all processes in comm instead
   simcall_process_kill(SIMIX_process_self());
   return MPI_SUCCESS;
@@ -127,7 +127,7 @@ int PMPI_Abort(MPI_Comm comm, int errorcode)
 double PMPI_Wtime(void)
 {
   double time;
-  if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_sample_is_running) {
+  if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_process_get_sampling()) {
     smpi_bench_end();
     time = SIMIX_get_clock();
     smpi_bench_begin();
@@ -147,14 +147,12 @@ int PMPI_Address(void *location, MPI_Aint * address)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (!address) {
     retval = MPI_ERR_ARG;
   } else {
     *address = (MPI_Aint) location;
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -167,14 +165,12 @@ int PMPI_Type_free(MPI_Datatype * datatype)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (!datatype) {
     retval = MPI_ERR_ARG;
   } else {
     smpi_datatype_free(datatype);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -182,7 +178,6 @@ int PMPI_Type_size(MPI_Datatype datatype, int *size)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (datatype == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (size == NULL) {
@@ -191,7 +186,6 @@ int PMPI_Type_size(MPI_Datatype datatype, int *size)
     *size = (int) smpi_datatype_size(datatype);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -199,7 +193,6 @@ int PMPI_Type_get_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (datatype == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (lb == NULL || extent == NULL) {
@@ -207,7 +200,6 @@ int PMPI_Type_get_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent
   } else {
     retval = smpi_datatype_extent(datatype, lb, extent);
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -220,7 +212,6 @@ int PMPI_Type_extent(MPI_Datatype datatype, MPI_Aint * extent)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (datatype == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (extent == NULL) {
@@ -229,7 +220,6 @@ int PMPI_Type_extent(MPI_Datatype datatype, MPI_Aint * extent)
     *extent = smpi_datatype_get_extent(datatype);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -237,7 +227,6 @@ int PMPI_Type_lb(MPI_Datatype datatype, MPI_Aint * disp)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (datatype == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (disp == NULL) {
@@ -246,7 +235,6 @@ int PMPI_Type_lb(MPI_Datatype datatype, MPI_Aint * disp)
     *disp = smpi_datatype_lb(datatype);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -254,7 +242,6 @@ int PMPI_Type_ub(MPI_Datatype datatype, MPI_Aint * disp)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (datatype == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (disp == NULL) {
@@ -263,7 +250,6 @@ int PMPI_Type_ub(MPI_Datatype datatype, MPI_Aint * disp)
     *disp = smpi_datatype_ub(datatype);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -271,14 +257,12 @@ int PMPI_Op_create(MPI_User_function * function, int commute, MPI_Op * op)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (function == NULL || op == NULL) {
     retval = MPI_ERR_ARG;
   } else {
     *op = smpi_op_new(function, commute);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -286,7 +270,6 @@ int PMPI_Op_free(MPI_Op * op)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (op == NULL) {
     retval = MPI_ERR_ARG;
   } else if (*op == MPI_OP_NULL) {
@@ -296,7 +279,6 @@ int PMPI_Op_free(MPI_Op * op)
     *op = MPI_OP_NULL;
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -304,7 +286,6 @@ int PMPI_Group_free(MPI_Group * group)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (group == NULL) {
     retval = MPI_ERR_ARG;
   } else {
@@ -312,7 +293,6 @@ int PMPI_Group_free(MPI_Group * group)
     *group = MPI_GROUP_NULL;
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -320,7 +300,6 @@ int PMPI_Group_size(MPI_Group group, int *size)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (group == MPI_GROUP_NULL) {
     retval = MPI_ERR_GROUP;
   } else if (size == NULL) {
@@ -329,7 +308,6 @@ int PMPI_Group_size(MPI_Group group, int *size)
     *size = smpi_group_size(group);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -337,7 +315,6 @@ int PMPI_Group_rank(MPI_Group group, int *rank)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (group == MPI_GROUP_NULL) {
     retval = MPI_ERR_GROUP;
   } else if (rank == NULL) {
@@ -346,7 +323,6 @@ int PMPI_Group_rank(MPI_Group group, int *rank)
     *rank = smpi_group_rank(group, smpi_process_index());
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -354,7 +330,6 @@ int PMPI_Group_translate_ranks(MPI_Group group1, int n, int *ranks1,
                               MPI_Group group2, int *ranks2)
 {
   int retval, i, index;
-  smpi_bench_end();
   if (group1 == MPI_GROUP_NULL || group2 == MPI_GROUP_NULL) {
     retval = MPI_ERR_GROUP;
   } else {
@@ -368,7 +343,6 @@ int PMPI_Group_translate_ranks(MPI_Group group1, int n, int *ranks1,
     }
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -376,7 +350,6 @@ int PMPI_Group_compare(MPI_Group group1, MPI_Group group2, int *result)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (group1 == MPI_GROUP_NULL || group2 == MPI_GROUP_NULL) {
     retval = MPI_ERR_GROUP;
   } else if (result == NULL) {
@@ -385,7 +358,6 @@ int PMPI_Group_compare(MPI_Group group1, MPI_Group group2, int *result)
     *result = smpi_group_compare(group1, group2);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -394,7 +366,6 @@ int PMPI_Group_union(MPI_Group group1, MPI_Group group2,
 {
   int retval, i, proc1, proc2, size, size2;
 
-  smpi_bench_end();
   if (group1 == MPI_GROUP_NULL || group2 == MPI_GROUP_NULL) {
     retval = MPI_ERR_GROUP;
   } else if (newgroup == NULL) {
@@ -425,7 +396,6 @@ int PMPI_Group_union(MPI_Group group1, MPI_Group group2,
     }
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -434,7 +404,6 @@ int PMPI_Group_intersection(MPI_Group group1, MPI_Group group2,
 {
   int retval, i, proc1, proc2, size;
 
-  smpi_bench_end();
   if (group1 == MPI_GROUP_NULL || group2 == MPI_GROUP_NULL) {
     retval = MPI_ERR_GROUP;
   } else if (newgroup == NULL) {
@@ -464,7 +433,6 @@ int PMPI_Group_intersection(MPI_Group group1, MPI_Group group2,
     }
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -472,7 +440,6 @@ int PMPI_Group_difference(MPI_Group group1, MPI_Group group2, MPI_Group * newgro
 {
   int retval, i, proc1, proc2, size, size2;
 
-  smpi_bench_end();
   if (group1 == MPI_GROUP_NULL || group2 == MPI_GROUP_NULL) {
     retval = MPI_ERR_GROUP;
   } else if (newgroup == NULL) {
@@ -500,7 +467,6 @@ int PMPI_Group_difference(MPI_Group group1, MPI_Group group2, MPI_Group * newgro
     }
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -508,7 +474,6 @@ int PMPI_Group_incl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
 {
   int retval, i, index;
 
-  smpi_bench_end();
   if (group == MPI_GROUP_NULL) {
     retval = MPI_ERR_GROUP;
   } else if (newgroup == NULL) {
@@ -532,7 +497,6 @@ int PMPI_Group_incl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
     }
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -540,7 +504,6 @@ int PMPI_Group_excl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
 {
   int retval, i, j, newsize, oldsize, index;
 
-  smpi_bench_end();
   if (group == MPI_GROUP_NULL) {
     retval = MPI_ERR_GROUP;
   } else if (newgroup == NULL) {
@@ -560,7 +523,7 @@ int PMPI_Group_excl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
       newsize = oldsize - n;
       *newgroup = smpi_group_new(newsize);
 
-      int* to_exclude=xbt_new(int, smpi_group_size(group));
+      int* to_exclude=xbt_new0(int, smpi_group_size(group));
       for(i=0; i<oldsize; i++)
         to_exclude[i]=0;
       for(i=0; i<n; i++)
@@ -579,7 +542,6 @@ int PMPI_Group_excl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
     }
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -588,7 +550,6 @@ int PMPI_Group_range_incl(MPI_Group group, int n, int ranges[][3],
 {
   int retval, i, j, rank, size, index;
 
-  smpi_bench_end();
   if (group == MPI_GROUP_NULL) {
     retval = MPI_ERR_GROUP;
   } else if (newgroup == NULL) {
@@ -637,7 +598,6 @@ int PMPI_Group_range_incl(MPI_Group group, int n, int ranges[][3],
     }
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -646,7 +606,6 @@ int PMPI_Group_range_excl(MPI_Group group, int n, int ranges[][3],
 {
   int retval, i, rank, newrank,oldrank, size, index, add;
 
-  smpi_bench_end();
   if (group == MPI_GROUP_NULL) {
     retval = MPI_ERR_GROUP;
   } else if (newgroup == NULL) {
@@ -715,15 +674,12 @@ int PMPI_Group_range_excl(MPI_Group group, int n, int ranges[][3],
 
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
 int PMPI_Comm_rank(MPI_Comm comm, int *rank)
 {
   int retval = 0;
-
-  smpi_bench_end();
   if (comm == MPI_COMM_NULL) {
     retval = MPI_ERR_COMM;
   } else if (rank == NULL) {
@@ -732,15 +688,12 @@ int PMPI_Comm_rank(MPI_Comm comm, int *rank)
     *rank = smpi_comm_rank(comm);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
 int PMPI_Comm_size(MPI_Comm comm, int *size)
 {
   int retval = 0;
-
-  smpi_bench_end();
   if (comm == MPI_COMM_NULL) {
     retval = MPI_ERR_COMM;
   } else if (size == NULL) {
@@ -749,7 +702,6 @@ int PMPI_Comm_size(MPI_Comm comm, int *size)
     *size = smpi_comm_size(comm);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -757,7 +709,6 @@ int PMPI_Comm_get_name (MPI_Comm comm, char* name, int* len)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (comm == MPI_COMM_NULL)  {
     retval = MPI_ERR_COMM;
   } else if (name == NULL || len == NULL)  {
@@ -766,7 +717,6 @@ int PMPI_Comm_get_name (MPI_Comm comm, char* name, int* len)
     smpi_comm_get_name(comm, name, len);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -774,7 +724,6 @@ int PMPI_Comm_group(MPI_Comm comm, MPI_Group * group)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (comm == MPI_COMM_NULL) {
     retval = MPI_ERR_COMM;
   } else if (group == NULL) {
@@ -788,7 +737,6 @@ int PMPI_Comm_group(MPI_Comm comm, MPI_Group * group)
     smpi_group_use(*group);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -796,7 +744,6 @@ int PMPI_Comm_compare(MPI_Comm comm1, MPI_Comm comm2, int *result)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (comm1 == MPI_COMM_NULL || comm2 == MPI_COMM_NULL) {
     retval = MPI_ERR_COMM;
   } else if (result == NULL) {
@@ -814,7 +761,6 @@ int PMPI_Comm_compare(MPI_Comm comm1, MPI_Comm comm2, int *result)
     }
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -822,7 +768,6 @@ int PMPI_Comm_dup(MPI_Comm comm, MPI_Comm * newcomm)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (comm == MPI_COMM_NULL) {
     retval = MPI_ERR_COMM;
   } else if (newcomm == NULL) {
@@ -831,7 +776,6 @@ int PMPI_Comm_dup(MPI_Comm comm, MPI_Comm * newcomm)
     *newcomm = smpi_comm_new(smpi_comm_group(comm));
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -839,7 +783,6 @@ int PMPI_Comm_create(MPI_Comm comm, MPI_Group group, MPI_Comm * newcomm)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (comm == MPI_COMM_NULL) {
     retval = MPI_ERR_COMM;
   } else if (group == MPI_GROUP_NULL) {
@@ -854,7 +797,6 @@ int PMPI_Comm_create(MPI_Comm comm, MPI_Group group, MPI_Comm * newcomm)
     *newcomm = smpi_comm_new(group);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -862,7 +804,6 @@ int PMPI_Comm_free(MPI_Comm * comm)
 {
   int retval = 0;
 
-  smpi_bench_end();
   if (comm == NULL) {
     retval = MPI_ERR_ARG;
   } else if (*comm == MPI_COMM_NULL) {
@@ -872,7 +813,6 @@ int PMPI_Comm_free(MPI_Comm * comm)
     *comm = MPI_COMM_NULL;
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -881,7 +821,6 @@ int PMPI_Comm_disconnect(MPI_Comm * comm)
   /* TODO: wait until all communication in comm are done */
   int retval = 0;
 
-  smpi_bench_end();
   if (comm == NULL) {
     retval = MPI_ERR_ARG;
   } else if (*comm == MPI_COMM_NULL) {
@@ -891,15 +830,14 @@ int PMPI_Comm_disconnect(MPI_Comm * comm)
     *comm = MPI_COMM_NULL;
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
 int PMPI_Comm_split(MPI_Comm comm, int color, int key, MPI_Comm* comm_out)
 {
   int retval = 0;
-
   smpi_bench_end();
+
   if (comm_out == NULL) {
     retval = MPI_ERR_ARG;
   } else if (comm == MPI_COMM_NULL) {
@@ -909,6 +847,7 @@ int PMPI_Comm_split(MPI_Comm comm, int color, int key, MPI_Comm* comm_out)
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
+
   return retval;
 }
 
@@ -1016,7 +955,6 @@ int PMPI_Request_free(MPI_Request * request)
   if (*request == MPI_REQUEST_NULL) {
     retval = MPI_ERR_ARG;
   } else {
-    if((*request)->flags & PERSISTENT)(*request)->refcount--;
     smpi_mpi_request_free(request);
     retval = MPI_SUCCESS;
   }
@@ -1053,7 +991,14 @@ int PMPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src,
 #ifdef HAVE_TRACING
     int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
     int src_traced = smpi_group_index(smpi_comm_group(comm), src);
-    TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__, count*smpi_datatype_size(datatype));
+
+    instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+    extra->type = TRACING_IRECV;
+    extra->send_size = count;
+    extra->src = src_traced;
+    extra->dst = rank;
+    extra->datatype1 = encode_datatype(datatype);
+    TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__, extra);
 #endif
 
     *request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
@@ -1099,9 +1044,15 @@ int PMPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst,
 
 #ifdef HAVE_TRACING
     int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-    TRACE_smpi_computing_out(rank);
     int dst_traced = smpi_group_index(smpi_comm_group(comm), dst);
-    TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, count*smpi_datatype_size(datatype));
+
+    instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+    extra->type = TRACING_ISEND;
+    extra->send_size = count;
+    extra->src = rank;
+    extra->dst = dst_traced;
+    extra->datatype1 = encode_datatype(datatype);
+    TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra);
     TRACE_smpi_send(rank, rank, dst_traced, count*smpi_datatype_size(datatype));
 #endif
 
@@ -1111,7 +1062,6 @@ int PMPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst,
 #ifdef HAVE_TRACING
     TRACE_smpi_ptp_out(rank, rank, dst_traced, __FUNCTION__);
     (*request)->send = 1;
-    TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -1148,9 +1098,14 @@ int PMPI_Issend(void* buf, int count, MPI_Datatype datatype,
 
 #ifdef HAVE_TRACING
     int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-    TRACE_smpi_computing_out(rank);
     int dst_traced = smpi_group_index(smpi_comm_group(comm), dst);
-    TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, count*smpi_datatype_size(datatype));
+    instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+    extra->type = TRACING_ISSEND;
+    extra->send_size = count;
+    extra->src = rank;
+    extra->dst = dst_traced;
+    extra->datatype1 = encode_datatype(datatype);
+    TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra);
     TRACE_smpi_send(rank, rank, dst_traced, count*smpi_datatype_size(datatype));
 #endif
 
@@ -1160,7 +1115,6 @@ int PMPI_Issend(void* buf, int count, MPI_Datatype datatype,
 #ifdef HAVE_TRACING
     TRACE_smpi_ptp_out(rank, rank, dst_traced, __FUNCTION__);
     (*request)->send = 1;
-    TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -1196,8 +1150,13 @@ int PMPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag,
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
   int src_traced = smpi_group_index(smpi_comm_group(comm), src);
-  TRACE_smpi_computing_out(rank);
-  TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__, count*smpi_datatype_size(datatype));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_RECV;
+  extra->send_size = count;
+  extra->src = src_traced;
+  extra->dst = rank;
+  extra->datatype1 = encode_datatype(datatype);
+  TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__, extra);
 #endif
 
     smpi_mpi_recv(buf, count, datatype, src, tag, comm, status);
@@ -1205,10 +1164,11 @@ int PMPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag,
 
 #ifdef HAVE_TRACING
   //the src may not have been known at the beginning of the recv (MPI_ANY_SOURCE)
-  if(status!=MPI_STATUS_IGNORE)src_traced = smpi_group_index(smpi_comm_group(comm), status->MPI_SOURCE);
+  if(status!=MPI_STATUS_IGNORE){
+    src_traced = smpi_group_index(smpi_comm_group(comm), status->MPI_SOURCE);
+    TRACE_smpi_recv(rank, src_traced, rank);
+  }
   TRACE_smpi_ptp_out(rank, src_traced, rank, __FUNCTION__);
-  TRACE_smpi_recv(rank, src_traced, rank);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -1241,9 +1201,14 @@ int PMPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag,
 
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
   int dst_traced = smpi_group_index(smpi_comm_group(comm), dst);
-  TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, count*smpi_datatype_size(datatype));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_SEND;
+  extra->send_size = count;
+  extra->src = rank;
+  extra->dst = dst_traced;
+  extra->datatype1 = encode_datatype(datatype);
+  TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra);
   TRACE_smpi_send(rank, rank, dst_traced,count*smpi_datatype_size(datatype));
 #endif
 
@@ -1252,7 +1217,6 @@ int PMPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag,
 
 #ifdef HAVE_TRACING
   TRACE_smpi_ptp_out(rank, rank, dst_traced, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -1285,10 +1249,14 @@ int PMPI_Ssend(void* buf, int count, MPI_Datatype datatype, int dst, int tag, MP
 
  #ifdef HAVE_TRACING
    int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-   TRACE_smpi_computing_out(rank);
    int dst_traced = smpi_group_index(smpi_comm_group(comm), dst);
-   TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, count*smpi_datatype_size(datatype));
-   TRACE_smpi_send(rank, rank, dst_traced,count*smpi_datatype_size(datatype));
+   instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+   extra->type = TRACING_SSEND;
+   extra->send_size = count;
+   extra->src = rank;
+   extra->dst = dst_traced;
+   extra->datatype1 = encode_datatype(datatype);
+   TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra);   TRACE_smpi_send(rank, rank, dst_traced,count*smpi_datatype_size(datatype));
  #endif
 
      smpi_mpi_ssend(buf, count, datatype, dst, tag, comm);
@@ -1296,7 +1264,6 @@ int PMPI_Ssend(void* buf, int count, MPI_Datatype datatype, int dst, int tag, MP
 
  #ifdef HAVE_TRACING
    TRACE_smpi_ptp_out(rank, rank, dst_traced, __FUNCTION__);
-   TRACE_smpi_computing_in(rank);
  #endif
    }
 
@@ -1335,10 +1302,18 @@ int PMPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
 
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
   int dst_traced = smpi_group_index(smpi_comm_group(comm), dst);
   int src_traced = smpi_group_index(smpi_comm_group(comm), src);
-  TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__, sendcount*smpi_datatype_size(sendtype));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_SENDRECV;
+  extra->send_size = sendcount;
+  extra->recv_size = recvcount;
+  extra->src = src_traced;
+  extra->dst = dst_traced;
+  extra->datatype1 = encode_datatype(sendtype);
+  extra->datatype2 = encode_datatype(recvtype);
+
+  TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__, extra);
   TRACE_smpi_send(rank, rank, dst_traced,sendcount*smpi_datatype_size(sendtype));
 #endif
 
@@ -1350,7 +1325,6 @@ int PMPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
 #ifdef HAVE_TRACING
   TRACE_smpi_ptp_out(rank, src_traced, dst_traced, __FUNCTION__);
   TRACE_smpi_recv(rank, src_traced, rank);
-  TRACE_smpi_computing_in(rank);
 #endif
 
   }
@@ -1372,7 +1346,7 @@ int PMPI_Sendrecv_replace(void *buf, int count, MPI_Datatype datatype,
       retval = MPI_ERR_COUNT;
   } else {
     int size = smpi_datatype_get_extent(datatype) * count;
-    recvbuf = xbt_new(char, size);
+    recvbuf = xbt_new0(char, size);
     retval =
         MPI_Sendrecv(buf, count, datatype, dst, sendtag, recvbuf, count,
                      datatype, src, recvtag, comm, status);
@@ -1497,13 +1471,14 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status)
     int rank = request && (*request)->comm != MPI_COMM_NULL
       ? smpi_process_index()
       : -1;
-    TRACE_smpi_computing_out(rank);
 
     int src_traced = (*request)->src;
     int dst_traced = (*request)->dst;
     MPI_Comm comm = (*request)->comm;
     int is_wait_for_receive = (*request)->recv;
-    TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__,-1);
+    instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+    extra->type = TRACING_WAIT;
+    TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__, extra);
 #endif
 
     smpi_mpi_wait(request, status);
@@ -1519,7 +1494,6 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status)
           src_traced;
       TRACE_smpi_recv(rank, src_traced, dst_traced);
     }
-    TRACE_smpi_computing_in(rank);
 #endif
 
   }
@@ -1530,16 +1504,17 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status)
 
 int PMPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status * status)
 {
-  int retval = 0;
+  if (index == NULL)
+    return MPI_ERR_ARG;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
   //save requests information for tracing
   int i;
-  int *srcs = xbt_new(int, count);
-  int *dsts = xbt_new(int, count);
-  int *recvs = xbt_new(int, count);
-  MPI_Comm *comms = xbt_new(MPI_Comm, count);
+  int *srcs = xbt_new0(int, count);
+  int *dsts = xbt_new0(int, count);
+  int *recvs = xbt_new0(int, count);
+  MPI_Comm *comms = xbt_new0(MPI_Comm, count);
 
   for (i = 0; i < count; i++) {
     MPI_Request req = requests[i];      //already received requests are no longer valid
@@ -1551,17 +1526,13 @@ int PMPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status * sta
     }
   }
   int rank_traced = smpi_process_index();
-  TRACE_smpi_computing_out(rank_traced);
-
-  TRACE_smpi_ptp_in(rank_traced, -1, -1, __FUNCTION__,count);
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_WAITANY;
+  extra->send_size=count;
+  TRACE_smpi_ptp_in(rank_traced, -1, -1, __FUNCTION__,extra);
 
 #endif
-  if (index == NULL) {
-    retval = MPI_ERR_ARG;
-  } else {
-    *index = smpi_mpi_waitany(count, requests, status);
-    retval = MPI_SUCCESS;
-  }
+  *index = smpi_mpi_waitany(count, requests, status);
 #ifdef HAVE_TRACING
   if(*index!=MPI_UNDEFINED){
     int src_traced = srcs[*index];
@@ -1582,10 +1553,9 @@ int PMPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status * sta
     xbt_free(comms);
 
   }
-  TRACE_smpi_computing_in(rank_traced);
 #endif
   smpi_bench_begin();
-  return retval;
+  return MPI_SUCCESS;
 }
 
 int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
@@ -1595,11 +1565,11 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
 #ifdef HAVE_TRACING
   //save information from requests
   int i;
-  int *srcs = xbt_new(int, count);
-  int *dsts = xbt_new(int, count);
-  int *recvs = xbt_new(int, count);
-  int *valid = xbt_new(int, count);
-  MPI_Comm *comms = xbt_new(MPI_Comm, count);
+  int *srcs = xbt_new0(int, count);
+  int *dsts = xbt_new0(int, count);
+  int *recvs = xbt_new0(int, count);
+  int *valid = xbt_new0(int, count);
+  MPI_Comm *comms = xbt_new0(MPI_Comm, count);
 
   //int valid_count = 0;
   for (i = 0; i < count; i++) {
@@ -1615,9 +1585,10 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
     }
   }
   int rank_traced = smpi_process_index();
-  TRACE_smpi_computing_out(rank_traced);
-
-  TRACE_smpi_ptp_in(rank_traced, -1, -1, __FUNCTION__,count);
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_WAITALL;
+  extra->send_size=count;
+  TRACE_smpi_ptp_in(rank_traced, -1, -1, __FUNCTION__,extra);
 #endif
   int retval = smpi_mpi_waitall(count, requests, status);
 #ifdef HAVE_TRACING
@@ -1644,7 +1615,6 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
   xbt_free(valid);
   xbt_free(comms);
 
-  TRACE_smpi_computing_in(rank_traced);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1694,15 +1664,20 @@ int PMPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm c
   } else {
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_index(smpi_comm_group(comm), root);
-  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__, count*smpi_datatype_size(datatype));
+
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_BCAST;
+  extra->send_size = count;
+  extra->root = root_traced;
+  extra->datatype1 = encode_datatype(datatype);
+  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__, extra);
+
 #endif
     mpi_coll_bcast_fun(buf, count, datatype, root, comm);
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -1721,14 +1696,14 @@ int PMPI_Barrier(MPI_Comm comm)
   } else {
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, smpi_comm_size(comm));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_BARRIER;
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, extra);
 #endif
     mpi_coll_barrier_fun(comm);
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -1763,9 +1738,16 @@ int PMPI_Gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
     }
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_index(smpi_comm_group(comm), root);
-  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__,sendcount*smpi_datatype_size(sendtmptype));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_GATHER;
+  extra->send_size = sendtmpcount;
+  extra->recv_size = recvcount;
+  extra->root = root_traced;
+  extra->datatype1 = encode_datatype(sendtmptype);
+  extra->datatype2 = encode_datatype(recvtype);
+
+  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__, extra);
 #endif
     mpi_coll_gather_fun(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcount,
                     recvtype, root, comm);
@@ -1774,7 +1756,6 @@ int PMPI_Gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -1810,16 +1791,27 @@ int PMPI_Gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
 
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_index(smpi_comm_group(comm), root);
-  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__,sendcount*smpi_datatype_size(sendtmptype));
+  int i=0;
+  int size = smpi_comm_size(comm);
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_GATHERV;
+  extra->send_size = sendtmpcount;
+  extra->recvcounts= xbt_malloc(size*sizeof(int));
+  for(i=0; i< size; i++)//copy data to avoid bad free
+    extra->recvcounts[i] = recvcounts[i];
+  extra->num_processes = size;
+  extra->root = root_traced;
+  extra->datatype1 = encode_datatype(sendtmptype);
+  extra->datatype2 = encode_datatype(recvtype);
+
+  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__,extra);
 #endif
     smpi_mpi_gatherv(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcounts,
                      displs, recvtype, root, comm);
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -1851,8 +1843,14 @@ int PMPI_Allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
     }
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,sendcount*smpi_datatype_size(sendtype));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_ALLGATHER;
+  extra->send_size = sendcount;
+  extra->recv_size = recvcount;
+  extra->datatype1 = encode_datatype(sendtype);
+  extra->datatype2 = encode_datatype(recvtype);
+
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, extra);
 #endif
     mpi_coll_allgather_fun(sendbuf, sendcount, sendtype, recvbuf, recvcount,
                            recvtype, comm);
@@ -1892,15 +1890,25 @@ int PMPI_Allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
     }
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,sendcount*smpi_datatype_size(sendtype));
+  int i=0;
+  int size = smpi_comm_size(comm);
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_ALLGATHERV;
+  extra->send_size = sendcount;
+  extra->recvcounts= xbt_malloc(size*sizeof(int));
+  for(i=0; i< size; i++)//copy data to avoid bad free
+    extra->recvcounts[i] = recvcounts[i];
+  extra->num_processes = size;
+  extra->datatype1 = encode_datatype(sendtype);
+  extra->datatype2 = encode_datatype(recvtype);
+
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 #endif
     mpi_coll_allgatherv_fun(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
                         displs, recvtype, comm);
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -1929,17 +1937,22 @@ int PMPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
     }
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_index(smpi_comm_group(comm), root);
-
-  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__,sendcount*smpi_datatype_size(recvtype));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_SCATTER;
+  extra->send_size = sendcount;
+  extra->recv_size= recvcount;
+  extra->root = root_traced;
+  extra->datatype1 = encode_datatype(sendtype);
+  extra->datatype2 = encode_datatype(recvtype);
+
+  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__,extra);
 #endif
     mpi_coll_scatter_fun(sendbuf, sendcount, sendtype, recvbuf, recvcount,
                      recvtype, root, comm);
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -1969,18 +1982,28 @@ int PMPI_Scatterv(void *sendbuf, int *sendcounts, int *displs,
     }
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_index(smpi_comm_group(comm), root);
-  int count=0, i;
-  for(i=0; i<smpi_comm_size(comm);i++)count+=sendcounts[i];
-  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__, count*smpi_datatype_size(sendtype));
+  int i=0;
+  int size = smpi_comm_size(comm);
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_SCATTERV;
+  extra->recv_size = recvcount;
+  extra->sendcounts= xbt_malloc(size*sizeof(int));
+  for(i=0; i< size; i++)//copy data to avoid bad free
+    extra->sendcounts[i] = sendcounts[i];
+  extra->num_processes = size;
+  extra->root = root_traced;
+  extra->datatype1 = encode_datatype(sendtype);
+  extra->datatype2 = encode_datatype(recvtype);
+
+  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__,extra);
+
 #endif
     smpi_mpi_scatterv(sendbuf, sendcounts, displs, sendtype, recvbuf,
                       recvcount, recvtype, root, comm);
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -2002,16 +2025,20 @@ int PMPI_Reduce(void *sendbuf, void *recvbuf, int count,
   } else {
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_index(smpi_comm_group(comm), root);
-  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__, count*smpi_datatype_size(datatype));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_REDUCE;
+  extra->send_size = count;
+  extra->datatype1 = encode_datatype(datatype);
+  extra->root = root_traced;
+
+  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__,extra);
 #endif
     mpi_coll_reduce_fun(sendbuf, recvbuf, count, datatype, op, root, comm);
 
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -2056,8 +2083,12 @@ int PMPI_Allreduce(void *sendbuf, void *recvbuf, int count,
     }
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, count*smpi_datatype_size(datatype));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_ALLREDUCE;
+  extra->send_size = count;
+  extra->datatype1 = encode_datatype(datatype);
+
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 #endif
     mpi_coll_allreduce_fun(sendtmpbuf, recvbuf, count, datatype, op, comm);
 
@@ -2068,7 +2099,6 @@ int PMPI_Allreduce(void *sendbuf, void *recvbuf, int count,
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -2092,14 +2122,17 @@ int PMPI_Scan(void *sendbuf, void *recvbuf, int count,
   } else {
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, count*smpi_datatype_size(datatype));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_SCAN;
+  extra->send_size = count;
+  extra->datatype1 = encode_datatype(datatype);
+
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 #endif
     smpi_mpi_scan(sendbuf, recvbuf, count, datatype, op, comm);
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -2122,14 +2155,17 @@ int PMPI_Exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
   } else {
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, count*smpi_datatype_size(datatype));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_EXSCAN;
+  extra->send_size = count;
+  extra->datatype1 = encode_datatype(datatype);
+
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 #endif
     smpi_mpi_exscan(sendbuf, recvbuf, count, datatype, op, comm);
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -2154,10 +2190,18 @@ int PMPI_Reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
   } else {
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
-  int count=0, i;
-  for(i=0; i<smpi_comm_size(comm);i++)count+=recvcounts[i];
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, count*smpi_datatype_size(datatype));
+  int i=0;
+  int size = smpi_comm_size(comm);
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_REDUCE_SCATTER;
+  extra->send_size = 0;
+  extra->recvcounts= xbt_malloc(size*sizeof(int));
+  for(i=0; i< size; i++)//copy data to avoid bad free
+    extra->recvcounts[i] = recvcounts[i];
+  extra->num_processes = size;
+  extra->datatype1 = encode_datatype(datatype);
+
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 #endif
     void* sendtmpbuf=sendbuf;
     if(sendbuf==MPI_IN_PLACE){
@@ -2169,7 +2213,6 @@ int PMPI_Reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -2192,12 +2235,21 @@ int PMPI_Reduce_scatter_block(void *sendbuf, void *recvbuf, int recvcount,
   } else if (recvcount < 0) {
     retval = MPI_ERR_ARG;
   } else {
+    int count=smpi_comm_size(comm);
+
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, recvcount*smpi_comm_size(comm)*smpi_datatype_size(datatype));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_REDUCE_SCATTER;
+  extra->send_size = 0;
+  extra->recvcounts= xbt_malloc(count*sizeof(int));
+  for(i=0; i< count; i++)//copy data to avoid bad free
+    extra->recvcounts[i] = recvcount;
+  extra->num_processes = count;
+  extra->datatype1 = encode_datatype(datatype);
+
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 #endif
-    int count=smpi_comm_size(comm);
     int* recvcounts=(int*)xbt_malloc(count);
     for (i=0; i<count;i++)recvcounts[i]=recvcount;
     mpi_coll_reduce_scatter_fun(sendbuf, recvbuf, recvcounts,
@@ -2206,7 +2258,6 @@ int PMPI_Reduce_scatter_block(void *sendbuf, void *recvbuf, int recvcount,
     retval = MPI_SUCCESS;
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -2230,13 +2281,18 @@ int PMPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   } else {
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, sendcount*smpi_datatype_size(sendtype));
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_ALLTOALL;
+  extra->send_size = sendcount;
+  extra->recv_size = recvcount;
+  extra->datatype1 = encode_datatype(sendtype);
+  extra->datatype2 = encode_datatype(recvtype);
+
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 #endif
     retval = mpi_coll_alltoall_fun(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -2263,10 +2319,28 @@ int PMPI_Alltoallv(void *sendbuf, int *sendcounts, int *senddisps,
   } else {
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  TRACE_smpi_computing_out(rank);
-  int i, size=0;
-  for(i=0; i< smpi_comm_size(comm);i++)size+=sendcounts[i];
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, size*smpi_datatype_size(sendtype));
+  int i=0;
+  int size = smpi_comm_size(comm);
+  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
+  extra->type = TRACING_ALLTOALLV;
+  extra->send_size = 0;
+  extra->recv_size = 0;
+  extra->recvcounts= xbt_malloc(size*sizeof(int));
+  extra->sendcounts= xbt_malloc(size*sizeof(int));
+
+  for(i=0; i< size; i++){//copy data to avoid bad free
+    extra->send_size += sendcounts[i];
+    extra->recv_size += recvcounts[i];
+
+    extra->sendcounts[i] = sendcounts[i];
+    extra->recvcounts[i] = recvcounts[i];
+  }
+  extra->num_processes = size;
+
+  extra->datatype1 = encode_datatype(sendtype);
+  extra->datatype2 = encode_datatype(recvtype);
+
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 #endif
     retval =
         mpi_coll_alltoallv_fun(sendbuf, sendcounts, senddisps, sendtype,
@@ -2274,7 +2348,6 @@ int PMPI_Alltoallv(void *sendbuf, int *sendcounts, int *senddisps,
                                   comm);
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
-  TRACE_smpi_computing_in(rank);
 #endif
   }
 
@@ -2287,7 +2360,6 @@ int PMPI_Get_processor_name(char *name, int *resultlen)
 {
   int retval = MPI_SUCCESS;
 
-  smpi_bench_end();
   strncpy(name, SIMIX_host_get_name(SIMIX_host_self()),
           strlen(SIMIX_host_get_name(SIMIX_host_self())) < MPI_MAX_PROCESSOR_NAME - 1 ?
           strlen(SIMIX_host_get_name(SIMIX_host_self())) +1 :
@@ -2296,7 +2368,6 @@ int PMPI_Get_processor_name(char *name, int *resultlen)
       strlen(name) >
       MPI_MAX_PROCESSOR_NAME ? MPI_MAX_PROCESSOR_NAME : strlen(name);
 
-  smpi_bench_begin();
   return retval;
 }
 
@@ -2305,7 +2376,6 @@ int PMPI_Get_count(MPI_Status * status, MPI_Datatype datatype, int *count)
   int retval = MPI_SUCCESS;
   size_t size;
 
-  smpi_bench_end();
   if (status == NULL || count == NULL) {
     retval = MPI_ERR_ARG;
   } else if (datatype == MPI_DATATYPE_NULL) {
@@ -2320,14 +2390,12 @@ int PMPI_Get_count(MPI_Status * status, MPI_Datatype datatype, int *count)
       *count = smpi_mpi_get_count(status, datatype);
     }
   }
-  smpi_bench_begin();
   return retval;
 }
 
 int PMPI_Type_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type) {
   int retval = 0;
 
-  smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (count<0){
@@ -2335,21 +2403,18 @@ int PMPI_Type_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_typ
   } else {
     retval = smpi_datatype_contiguous(count, old_type, new_type, 0);
   }
-  smpi_bench_begin();
   return retval;
 }
 
 int PMPI_Type_commit(MPI_Datatype* datatype) {
   int retval = 0;
 
-  smpi_bench_end();
   if (datatype == NULL || *datatype == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else {
     smpi_datatype_commit(datatype);
     retval = MPI_SUCCESS;
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -2357,7 +2422,6 @@ int PMPI_Type_commit(MPI_Datatype* datatype) {
 int PMPI_Type_vector(int count, int blocklen, int stride, MPI_Datatype old_type, MPI_Datatype* new_type) {
   int retval = 0;
 
-  smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (count<0 || blocklen<0){
@@ -2365,14 +2429,12 @@ int PMPI_Type_vector(int count, int blocklen, int stride, MPI_Datatype old_type,
   } else {
     retval = smpi_datatype_vector(count, blocklen, stride, old_type, new_type);
   }
-  smpi_bench_begin();
   return retval;
 }
 
 int PMPI_Type_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type) {
   int retval = 0;
 
-  smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (count<0 || blocklen<0){
@@ -2380,7 +2442,6 @@ int PMPI_Type_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old
   } else {
     retval = smpi_datatype_hvector(count, blocklen, stride, old_type, new_type);
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -2391,7 +2452,6 @@ int PMPI_Type_create_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datat
 int PMPI_Type_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
   int retval = 0;
 
-  smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (count<0){
@@ -2399,14 +2459,12 @@ int PMPI_Type_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_
   } else {
     retval = smpi_datatype_indexed(count, blocklens, indices, old_type, new_type);
   }
-  smpi_bench_begin();
   return retval;
 }
 
 int PMPI_Type_create_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
   int retval = 0;
 
-  smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (count<0){
@@ -2414,14 +2472,12 @@ int PMPI_Type_create_indexed(int count, int* blocklens, int* indices, MPI_Dataty
   } else {
     retval = smpi_datatype_indexed(count, blocklens, indices, old_type, new_type);
   }
-  smpi_bench_begin();
   return retval;
 }
 
 int PMPI_Type_create_indexed_block(int count, int blocklength, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
   int retval,i;
 
-  smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (count<0){
@@ -2432,7 +2488,6 @@ int PMPI_Type_create_indexed_block(int count, int blocklength, int* indices, MPI
     retval = smpi_datatype_indexed(count, blocklens, indices, old_type, new_type);
     xbt_free(blocklens);
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -2440,7 +2495,6 @@ int PMPI_Type_create_indexed_block(int count, int blocklength, int* indices, MPI
 int PMPI_Type_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
   int retval = 0;
 
-  smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (count<0){
@@ -2448,7 +2502,6 @@ int PMPI_Type_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatyp
   } else {
     retval = smpi_datatype_hindexed(count, blocklens, indices, old_type, new_type);
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -2459,7 +2512,6 @@ int PMPI_Type_create_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_
 int PMPI_Type_create_hindexed_block(int count, int blocklength, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
   int retval,i;
 
-  smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else if (count<0){
@@ -2470,7 +2522,6 @@ int PMPI_Type_create_hindexed_block(int count, int blocklength, MPI_Aint* indice
     retval = smpi_datatype_hindexed(count, blocklens, indices, old_type, new_type);
     xbt_free(blocklens);
   }
-  smpi_bench_begin();
   return retval;
 }
 
@@ -2478,13 +2529,11 @@ int PMPI_Type_create_hindexed_block(int count, int blocklength, MPI_Aint* indice
 int PMPI_Type_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type) {
   int retval = 0;
 
-  smpi_bench_end();
   if (count<0){
     retval = MPI_ERR_COUNT;
   } else {
     retval = smpi_datatype_struct(count, blocklens, indices, old_types, new_type);
   }
-  smpi_bench_begin();
   return retval;
 }