add option to generate states for code outside smpi to allow computation timing ...

diff --git a/src/instr/instr_config.c b/src/instr/instr_config.c
index 40ad2bf..16065f1 100644 (file)
--- a/src/instr/instr_config.c
+++ b/src/instr/instr_config.c
@@ -16,6 +16,7 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY (instr_config, instr, "Configuration");
 #define OPT_TRACING_PLATFORM      "tracing/platform"
 #define OPT_TRACING_SMPI          "tracing/smpi"
 #define OPT_TRACING_SMPI_GROUP    "tracing/smpi/group"
+#define OPT_TRACING_SMPI_COMPUTING "tracing/smpi/computing"
 #define OPT_TRACING_CATEGORIZED   "tracing/categorized"
 #define OPT_TRACING_UNCATEGORIZED "tracing/uncategorized"
 #define OPT_TRACING_MSG_PROCESS   "tracing/msg/process"
@@ -33,6 +34,7 @@ static int trace_enabled;
 static int trace_platform;
 static int trace_smpi_enabled;
 static int trace_smpi_grouped;
+static int trace_smpi_computing;
 static int trace_categorized;
 static int trace_uncategorized;
 static int trace_msg_process_enabled;
@@ -50,6 +52,7 @@ static void TRACE_getopts(void)
   trace_platform = xbt_cfg_get_int(_surf_cfg_set, OPT_TRACING_PLATFORM);
   trace_smpi_enabled = xbt_cfg_get_int(_surf_cfg_set, OPT_TRACING_SMPI);
   trace_smpi_grouped = xbt_cfg_get_int(_surf_cfg_set, OPT_TRACING_SMPI_GROUP);
+  trace_smpi_computing = xbt_cfg_get_int(_surf_cfg_set, OPT_TRACING_SMPI_COMPUTING);
   trace_categorized = xbt_cfg_get_int(_surf_cfg_set, OPT_TRACING_CATEGORIZED);
   trace_uncategorized = xbt_cfg_get_int(_surf_cfg_set, OPT_TRACING_UNCATEGORIZED);
   trace_msg_process_enabled = xbt_cfg_get_int(_surf_cfg_set, OPT_TRACING_MSG_PROCESS);
@@ -163,6 +166,12 @@ int TRACE_smpi_is_grouped(void)
   return trace_smpi_grouped;
 }
 
+int TRACE_smpi_is_computing(void)
+{
+  return trace_smpi_computing;
+}
+
+
 int TRACE_categorized (void)
 {
   return trace_categorized;
@@ -261,6 +270,12 @@ void TRACE_global_init(int *argc, char **argv)
                    xbt_cfgelm_int, &default_tracing_smpi_grouped, 0, 1,
                    NULL, NULL);
 
+  /* smpi computing */
+  int default_tracing_smpi_computing = 0;
+  xbt_cfg_register(&_surf_cfg_set, OPT_TRACING_SMPI_COMPUTING,
+                   "Generate states for timing out of SMPI parts of the application",
+                   xbt_cfgelm_int, &default_tracing_smpi_computing, 0, 1,
+                   NULL, NULL);
 
   /* tracing categorized resource utilization traces */
   int default_tracing_categorized = 0;
@@ -389,6 +404,10 @@ void TRACE_help (int detailed)
       "  This option only has effect if this simulator is SMPI-based. The processes\n"
       "  are grouped by the hosts where they were executed.",
       detailed);
+  print_line (OPT_TRACING_SMPI_COMPUTING, "Generates a \" Computing \" State",
+         "  This option aims at tracing computations in the application, outside SMPI\n"
+      "  to allow further study of simulated or real computation time",
+      detailed);
   print_line (OPT_TRACING_MSG_PROCESS, "Trace processes behavior (MSG)",
       "  This option only has effect if this simulator is MSG-based. It traces the\n"
       "  behavior of all categorized MSG processes, grouping them by hosts. This option\n"

diff --git a/src/instr/instr_private.h b/src/instr/instr_private.h
index 58b4475..27e58ca 100644 (file)
--- a/src/instr/instr_private.h
+++ b/src/instr/instr_private.h
@@ -171,6 +171,9 @@ void TRACE_smpi_init(int rank);
 void TRACE_smpi_finalize(int rank);
 void TRACE_smpi_collective_in(int rank, int root, const char *operation);
 void TRACE_smpi_collective_out(int rank, int root, const char *operation);
+void TRACE_smpi_computing_init(int rank);
+void TRACE_smpi_computing_out(int rank);
+void TRACE_smpi_computing_in(int rank);
 void TRACE_smpi_ptp_in(int rank, int src, int dst, const char *operation);
 void TRACE_smpi_ptp_out(int rank, int src, int dst, const char *operation);
 void TRACE_smpi_send(int rank, int src, int dst);
@@ -185,6 +188,7 @@ int TRACE_platform(void);
 int TRACE_is_configured(void);
 int TRACE_smpi_is_enabled(void);
 int TRACE_smpi_is_grouped(void);
+int TRACE_smpi_is_computing(void);
 int TRACE_categorized (void);
 int TRACE_uncategorized (void);
 int TRACE_msg_process_is_enabled(void);

diff --git a/src/instr/instr_smpi.c b/src/instr/instr_smpi.c
index 30f1802..a5247af 100644 (file)
--- a/src/instr/instr_smpi.c
+++ b/src/instr/instr_smpi.c
@@ -200,6 +200,45 @@ void TRACE_smpi_collective_out(int rank, int root, const char *operation)
   new_pajePopState (SIMIX_get_clock(), container, type);
 }
 
+void TRACE_smpi_computing_init(int rank)
+{
+ //first use, initialize the color in the trace
+ //TODO : check with lucas and Pierre how to generalize this approach
+  //to avoid unnecessary access to the color array
+  if (!TRACE_smpi_is_enabled() || !TRACE_smpi_is_computing()) return;
+
+  char str[INSTR_DEFAULT_STR_SIZE];
+  smpi_container(rank, str, INSTR_DEFAULT_STR_SIZE);
+  container_t container = PJ_container_get (str);
+  type_t type = PJ_type_get ("MPI_STATE", container->type);
+  const char *color = instr_find_color ("computing");
+  val_t value = PJ_value_get_or_new ("computing", color, type);
+  new_pajePushState (SIMIX_get_clock(), container, type, value);
+}
+
+void TRACE_smpi_computing_in(int rank)
+{
+  //do not forget to set the color first, otherwise this will explode
+  if (!TRACE_smpi_is_enabled()|| !TRACE_smpi_is_computing()) return;
+
+  char str[INSTR_DEFAULT_STR_SIZE];
+  smpi_container(rank, str, INSTR_DEFAULT_STR_SIZE);
+  container_t container = PJ_container_get (str);
+  type_t type = PJ_type_get ("MPI_STATE", container->type);
+  val_t value = PJ_value_get_or_new ("computing", NULL, type);
+  new_pajePushState (SIMIX_get_clock(), container, type, value);
+}
+
+void TRACE_smpi_computing_out(int rank)
+{
+  if (!TRACE_smpi_is_enabled()|| !TRACE_smpi_is_computing()) return;
+  char str[INSTR_DEFAULT_STR_SIZE];
+  smpi_container(rank, str, INSTR_DEFAULT_STR_SIZE);
+  container_t container = PJ_container_get (str);
+  type_t type = PJ_type_get ("MPI_STATE", container->type);
+  new_pajePopState (SIMIX_get_clock(), container, type);
+}
+
 void TRACE_smpi_ptp_in(int rank, int src, int dst, const char *operation)
 {
   if (!TRACE_smpi_is_enabled()) return;

diff --git a/src/smpi/smpi_pmpi.c b/src/smpi/smpi_pmpi.c
index db8815d..54464cb 100644 (file)
--- a/src/smpi/smpi_pmpi.c
+++ b/src/smpi/smpi_pmpi.c
@@ -28,7 +28,10 @@ int PMPI_Init(int *argc, char ***argv)
 {
   smpi_process_init(argc, argv);
 #ifdef HAVE_TRACING
-  TRACE_smpi_init(smpi_process_index());
+  int rank = smpi_process_index();
+  TRACE_smpi_init(rank);
+
+  TRACE_smpi_computing_init(rank);
 #endif
   smpi_bench_begin();
   return MPI_SUCCESS;
@@ -39,6 +42,8 @@ int PMPI_Finalize(void)
   smpi_process_finalize();
   smpi_bench_end();
 #ifdef HAVE_TRACING
+  int rank = smpi_process_index();
+  TRACE_smpi_computing_out(rank);
   TRACE_smpi_finalize(smpi_process_index());
 #endif
   smpi_process_destroy();
@@ -87,6 +92,10 @@ 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;
@@ -916,6 +925,7 @@ int PMPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   int dst_traced = smpi_group_rank(smpi_comm_group(comm), dst);
   TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__);
   TRACE_smpi_send(rank, rank, dst_traced);
@@ -931,6 +941,7 @@ 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
   smpi_bench_begin();
   return retval;
@@ -945,6 +956,8 @@ int PMPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag,
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
   int src_traced = smpi_group_rank(smpi_comm_group(comm), src);
+  TRACE_smpi_computing_out(rank);
+
   TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__);
 #endif
   if (comm == MPI_COMM_NULL) {
@@ -956,6 +969,7 @@ int PMPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag,
 #ifdef HAVE_TRACING
   TRACE_smpi_ptp_out(rank, src_traced, rank, __FUNCTION__);
   TRACE_smpi_recv(rank, src_traced, rank);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -969,6 +983,7 @@ int PMPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   int dst_traced = smpi_group_rank(smpi_comm_group(comm), dst);
   TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__);
   TRACE_smpi_send(rank, rank, dst_traced);
@@ -981,6 +996,7 @@ 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
   smpi_bench_begin();
   return retval;
@@ -996,6 +1012,7 @@ int PMPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   int dst_traced = smpi_group_rank(smpi_comm_group(comm), dst);
   int src_traced = smpi_group_rank(smpi_comm_group(comm), src);
   TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__);
@@ -1016,6 +1033,8 @@ int PMPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   TRACE_smpi_ptp_out(rank, src_traced, dst_traced, __FUNCTION__);
   TRACE_smpi_recv(rank, rank, dst_traced);
   TRACE_smpi_recv(rank, src_traced, rank);
+  TRACE_smpi_computing_in(rank);
+
 #endif
   smpi_bench_begin();
   return retval;
@@ -1081,6 +1100,8 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status)
   int rank = request && (*request)->comm != MPI_COMM_NULL
       ? smpi_comm_rank((*request)->comm)
       : -1;
+  TRACE_smpi_computing_out(rank);
+
   MPI_Group group = smpi_comm_group((*request)->comm);
   int src_traced = smpi_group_rank(group, (*request)->src);
   int dst_traced = smpi_group_rank(group, (*request)->dst);
@@ -1100,6 +1121,8 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status)
   if (is_wait_for_receive) {
     TRACE_smpi_recv(rank, src_traced, dst_traced);
   }
+  TRACE_smpi_computing_in(rank);
+
 #endif
   smpi_bench_begin();
   return retval;
@@ -1140,7 +1163,10 @@ int PMPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status * sta
     }
   }
   int rank_traced = smpi_comm_rank(MPI_COMM_WORLD);
+  TRACE_smpi_computing_out(rank_traced);
+
   TRACE_smpi_ptp_in(rank_traced, -1, -1, __FUNCTION__);
+
 #endif
   if (index == NULL) {
     retval = MPI_ERR_ARG;
@@ -1161,6 +1187,8 @@ int PMPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status * sta
   xbt_dynar_free(&srcs);
   xbt_dynar_free(&dsts);
   xbt_dynar_free(&recvs);
+  TRACE_smpi_computing_in(rank_traced);
+
 #endif
   smpi_bench_begin();
   return retval;
@@ -1192,6 +1220,8 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
     xbt_free(arecv);
   }
   int rank_traced = smpi_comm_rank (MPI_COMM_WORLD);
+  TRACE_smpi_computing_out(rank_traced);
+
   TRACE_smpi_ptp_in(rank_traced, -1, -1, __FUNCTION__);
 #endif
   smpi_mpi_waitall(count, requests, status);
@@ -1210,6 +1240,7 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
   xbt_dynar_free(&srcs);
   xbt_dynar_free(&dsts);
   xbt_dynar_free(&recvs);
+  TRACE_smpi_computing_in(rank_traced);
 #endif
   smpi_bench_begin();
   return MPI_SUCCESS;
@@ -1238,6 +1269,7 @@ int PMPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm c
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_rank(smpi_comm_group(comm), root);
   TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__);
 #endif
@@ -1249,6 +1281,7 @@ int PMPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm c
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1261,6 +1294,7 @@ int PMPI_Barrier(MPI_Comm comm)
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
 #endif
   if (comm == MPI_COMM_NULL) {
@@ -1271,6 +1305,7 @@ int PMPI_Barrier(MPI_Comm comm)
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1285,6 +1320,7 @@ int PMPI_Gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_rank(smpi_comm_group(comm), root);
   TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__);
 #endif
@@ -1300,6 +1336,7 @@ int PMPI_Gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1314,6 +1351,7 @@ int PMPI_Gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_rank(smpi_comm_group(comm), root);
   TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__);
 #endif
@@ -1331,6 +1369,7 @@ int PMPI_Gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1345,6 +1384,7 @@ int PMPI_Allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
 #endif
   if (comm == MPI_COMM_NULL) {
@@ -1373,6 +1413,7 @@ int PMPI_Allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
 #endif
   if (comm == MPI_COMM_NULL) {
@@ -1389,6 +1430,7 @@ int PMPI_Allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1403,6 +1445,7 @@ int PMPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_rank(smpi_comm_group(comm), root);
   TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__);
 #endif
@@ -1418,6 +1461,7 @@ int PMPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1432,6 +1476,7 @@ int PMPI_Scatterv(void *sendbuf, int *sendcounts, int *displs,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_rank(smpi_comm_group(comm), root);
   TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__);
 #endif
@@ -1449,6 +1494,7 @@ int PMPI_Scatterv(void *sendbuf, int *sendcounts, int *displs,
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1462,6 +1508,7 @@ int PMPI_Reduce(void *sendbuf, void *recvbuf, int count,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_rank(smpi_comm_group(comm), root);
   TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__);
 #endif
@@ -1475,6 +1522,7 @@ int PMPI_Reduce(void *sendbuf, void *recvbuf, int count,
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1488,6 +1536,7 @@ int PMPI_Allreduce(void *sendbuf, void *recvbuf, int count,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
 #endif
   if (comm == MPI_COMM_NULL) {
@@ -1502,6 +1551,7 @@ int PMPI_Allreduce(void *sendbuf, void *recvbuf, int count,
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1515,6 +1565,7 @@ int PMPI_Scan(void *sendbuf, void *recvbuf, int count,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
 #endif
   if (comm == MPI_COMM_NULL) {
@@ -1529,6 +1580,7 @@ int PMPI_Scan(void *sendbuf, void *recvbuf, int count,
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1543,6 +1595,7 @@ int PMPI_Reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
+  TRACE_smpi_computing_out(rank);
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
 #endif
   if (comm == MPI_COMM_NULL) {
@@ -1571,6 +1624,7 @@ int PMPI_Reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1585,6 +1639,7 @@ int PMPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
 #endif
   if (comm == MPI_COMM_NULL) {
@@ -1614,6 +1669,7 @@ int PMPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;
@@ -1628,6 +1684,7 @@ int PMPI_Alltoallv(void *sendbuf, int *sendcounts, int *senddisps,
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+  TRACE_smpi_computing_out(rank);
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
 #endif
   if (comm == MPI_COMM_NULL) {
@@ -1646,6 +1703,7 @@ int PMPI_Alltoallv(void *sendbuf, int *sendcounts, int *senddisps,
   }
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
 #endif
   smpi_bench_begin();
   return retval;