X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/fbb6b4556c347a1dec9e462574fdc2967f494002..b784c77ac5599c578f0ff00d6102222b286b0833:/src/smpi/smpi_replay.c

diff --git a/src/smpi/smpi_replay.c b/src/smpi/smpi_replay.c
index d7990bf2ab..fbfb10709d 100644
--- a/src/smpi/smpi_replay.c
+++ b/src/smpi/smpi_replay.c
@@ -1,4 +1,4 @@
-/* Copyright (c) 2009, 2010, 2011, 2012. The SimGrid Team.
+/* Copyright (c) 2009-2013. The SimGrid Team.
  * All rights reserved.                                                     */
 
 /* This program is free software; you can redistribute it and/or modify it
@@ -14,9 +14,10 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_replay,smpi,"Trace Replay with SMPI");
 
 int communicator_size = 0;
 static int active_processes = 0;
-xbt_dynar_t *reqq;
+xbt_dynar_t *reqq = NULL;
 
-MPI_Datatype MPI_DEFAULT_TYPE, MPI_CURRENT_TYPE;
+MPI_Datatype MPI_DEFAULT_TYPE;
+MPI_Datatype MPI_CURRENT_TYPE;
 
 static void log_timed_action (const char *const *action, double clock){
   if (XBT_LOG_ISENABLED(smpi_replay, xbt_log_priority_verbose)){
@@ -27,7 +28,6 @@ static void log_timed_action (const char *const *action, double clock){
 }
 
 typedef struct {
-  xbt_dynar_t isends; /* of MPI_Request */
   xbt_dynar_t irecvs; /* of MPI_Request */
 } s_smpi_replay_globals_t, *smpi_replay_globals_t;
 
@@ -46,7 +46,7 @@ static double parse_double(const char *string)
 static MPI_Datatype decode_datatype(const char *const action)
 {
 // Declared datatypes,
- 
+
   switch(atoi(action))
   {
     case 0:
@@ -72,7 +72,7 @@ static MPI_Datatype decode_datatype(const char *const action)
       break;
     default:
       MPI_CURRENT_TYPE=MPI_DEFAULT_TYPE;
-  
+
   }
    return MPI_CURRENT_TYPE;
 }
@@ -82,12 +82,11 @@ static void action_init(const char *const *action)
   int i;
   XBT_DEBUG("Initialize the counters");
   smpi_replay_globals_t globals =  xbt_new(s_smpi_replay_globals_t, 1);
-  globals->isends = xbt_dynar_new(sizeof(MPI_Request),NULL);
   globals->irecvs = xbt_dynar_new(sizeof(MPI_Request),NULL);
 
   if(action[2]) MPI_DEFAULT_TYPE= MPI_DOUBLE; // default MPE dataype 
   else MPI_DEFAULT_TYPE= MPI_BYTE; // default TAU datatype
-  
+
   smpi_process_set_user_data((void*) globals);
 
   /* start a simulated timer */
@@ -95,10 +94,12 @@ static void action_init(const char *const *action)
   /*initialize the number of active processes */
   active_processes = smpi_process_count();
 
-  reqq=xbt_new0(xbt_dynar_t,active_processes);
-  
-  for(i=0;i<active_processes;i++){
-    reqq[i]=xbt_dynar_new(sizeof(MPI_Request),NULL);
+  if (!reqq) {
+    reqq=xbt_new0(xbt_dynar_t,active_processes);
+
+    for(i=0;i<active_processes;i++){
+      reqq[i]=xbt_dynar_new(sizeof(MPI_Request),NULL);
+    }
   }
 }
 
@@ -107,9 +108,8 @@ static void action_finalize(const char *const *action)
   smpi_replay_globals_t globals =
       (smpi_replay_globals_t) smpi_process_get_user_data();
   if (globals){
-    XBT_DEBUG("There are %lu isends and %lu irecvs in the dynars",
-         xbt_dynar_length(globals->isends),xbt_dynar_length(globals->irecvs));
-    xbt_dynar_free_container(&(globals->isends));
+    XBT_DEBUG("There are %lu irecvs in the dynar",
+         xbt_dynar_length(globals->irecvs));
     xbt_dynar_free_container(&(globals->irecvs));
   }
   free(globals);
@@ -156,13 +156,13 @@ static void action_send(const char *const *action)
   } else {
     MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE;
   }
-    
+
 #ifdef HAVE_TRACING
   int rank = smpi_comm_rank(MPI_COMM_WORLD);
   TRACE_smpi_computing_out(rank);
   int dst_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), to);
-  TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__);
-  TRACE_smpi_send(rank, rank, dst_traced);
+  TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, size*smpi_datatype_size(MPI_CURRENT_TYPE));
+  TRACE_smpi_send(rank, rank, dst_traced, size*smpi_datatype_size(MPI_CURRENT_TYPE));
 #endif
 
   smpi_mpi_send(NULL, size, MPI_CURRENT_TYPE, to , 0, MPI_COMM_WORLD);
@@ -186,25 +186,22 @@ static void action_Isend(const char *const *action)
   if(action[4]) MPI_CURRENT_TYPE=decode_datatype(action[4]);
   else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE;
 
-  smpi_replay_globals_t globals =
-     (smpi_replay_globals_t) smpi_process_get_user_data();
 #ifdef HAVE_TRACING
   int rank = smpi_comm_rank(MPI_COMM_WORLD);
   TRACE_smpi_computing_out(rank);
   int dst_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), to);
-  TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__);
-  TRACE_smpi_send(rank, rank, dst_traced);
+  TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, size*smpi_datatype_size(MPI_CURRENT_TYPE));
+  TRACE_smpi_send(rank, rank, dst_traced, size*smpi_datatype_size(MPI_CURRENT_TYPE));
 #endif
 
   request = smpi_mpi_isend(NULL, size, MPI_CURRENT_TYPE, to, 0,MPI_COMM_WORLD);
-  
+
 #ifdef HAVE_TRACING
   TRACE_smpi_ptp_out(rank, rank, dst_traced, __FUNCTION__);
   request->send = 1;
   TRACE_smpi_computing_in(rank);
 #endif
 
-  xbt_dynar_push(globals->isends,&request);
   xbt_dynar_push(reqq[smpi_comm_rank(MPI_COMM_WORLD)],&request);
 
   log_timed_action (action, clock);
@@ -218,13 +215,13 @@ static void action_recv(const char *const *action) {
 
   if(action[4]) MPI_CURRENT_TYPE=decode_datatype(action[4]);
   else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE;
-  
+
 #ifdef HAVE_TRACING
   int rank = smpi_comm_rank(MPI_COMM_WORLD);
   int src_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), from);
   TRACE_smpi_computing_out(rank);
 
-  TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__);
+  TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__, size*smpi_datatype_size(MPI_CURRENT_TYPE));
 #endif
 
   smpi_mpi_recv(NULL, size, MPI_CURRENT_TYPE, from, 0, MPI_COMM_WORLD, &status);
@@ -247,18 +244,18 @@ static void action_Irecv(const char *const *action)
 
   smpi_replay_globals_t globals =
      (smpi_replay_globals_t) smpi_process_get_user_data();
-  
+
   if(action[4]) MPI_CURRENT_TYPE=decode_datatype(action[4]);
   else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE;
 
 #ifdef HAVE_TRACING
   int rank = smpi_comm_rank(MPI_COMM_WORLD);
   int src_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), from);
-  TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__);
+  TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__, size*smpi_datatype_size(MPI_CURRENT_TYPE));
 #endif
 
   request = smpi_mpi_irecv(NULL, size, MPI_CURRENT_TYPE, from, 0, MPI_COMM_WORLD);
-  
+
 #ifdef HAVE_TRACING
   TRACE_smpi_ptp_out(rank, src_traced, rank, __FUNCTION__);
   request->recv = 1;
@@ -290,7 +287,7 @@ static void action_wait(const char *const *action){
   int src_traced = smpi_group_rank(group, request->src);
   int dst_traced = smpi_group_rank(group, request->dst);
   int is_wait_for_receive = request->recv;
-  TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__);
+  TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__, -1);
 #endif
   smpi_mpi_wait(&request, &status);
 #ifdef HAVE_TRACING
@@ -314,14 +311,14 @@ static void action_waitall(const char *const *action){
   if (count_requests>0) {
     MPI_Request requests[count_requests];
     MPI_Status status[count_requests];
-  
+
     /*  The reqq is an array of dynars. Its index corresponds to the rank.
      Thus each rank saves its own requests to the array request. */
     xbt_dynar_foreach(reqq[smpi_comm_rank(MPI_COMM_WORLD)],i,requests[i]); 
-    
+
   #ifdef HAVE_TRACING
    //save information from requests
- 
+
    xbt_dynar_t srcs = xbt_dynar_new(sizeof(int), NULL);
    xbt_dynar_t dsts = xbt_dynar_new(sizeof(int), NULL);
    xbt_dynar_t recvs = xbt_dynar_new(sizeof(int), NULL);
@@ -350,7 +347,7 @@ static void action_waitall(const char *const *action){
    int rank_traced = smpi_process_index();
    TRACE_smpi_computing_out(rank_traced);
 
-   TRACE_smpi_ptp_in(rank_traced, -1, -1, __FUNCTION__);
+   TRACE_smpi_ptp_in(rank_traced, -1, -1, __FUNCTION__, count_requests);
   #endif
 
     smpi_mpi_waitall(count_requests, requests, status);
@@ -372,7 +369,7 @@ static void action_waitall(const char *const *action){
    xbt_dynar_free(&recvs);
    TRACE_smpi_computing_in(rank_traced);
   #endif
-   
+
    xbt_dynar_reset(reqq[smpi_comm_rank(MPI_COMM_WORLD)]);
   }
   log_timed_action (action, clock);
@@ -383,7 +380,7 @@ static void action_barrier(const char *const *action){
 #ifdef HAVE_TRACING
   int rank = smpi_comm_rank(MPI_COMM_WORLD);
   TRACE_smpi_computing_out(rank);
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, smpi_comm_size(MPI_COMM_WORLD));
 #endif
   smpi_mpi_barrier(MPI_COMM_WORLD);
 #ifdef HAVE_TRACING
@@ -412,12 +409,12 @@ static void action_bcast(const char *const *action)
       MPI_CURRENT_TYPE=decode_datatype(action[4]);   
     }
   }
-  
+
 #ifdef HAVE_TRACING
   int rank = smpi_comm_rank(MPI_COMM_WORLD);
   TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), 0);
-  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__);
+  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__,size*smpi_datatype_size(MPI_CURRENT_TYPE));
 #endif
 
   smpi_mpi_bcast(NULL, size, MPI_CURRENT_TYPE, root, MPI_COMM_WORLD);
@@ -436,21 +433,21 @@ static void action_reduce(const char *const *action)
   double clock = smpi_process_simulated_elapsed();
   int root=0;
   MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE;
-  
+
   if(action[4]) {
-      root= atoi(action[4]);
-      if(action[5]) {
-	MPI_CURRENT_TYPE=decode_datatype(action[5]);
-      }
+    root= atoi(action[4]);
+    if(action[5]) {
+      MPI_CURRENT_TYPE=decode_datatype(action[5]);
+    }
   }
 
 #ifdef HAVE_TRACING
   int rank = smpi_comm_rank(MPI_COMM_WORLD);
   TRACE_smpi_computing_out(rank);
   int root_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), 0);
-  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__);
+  TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__,comm_size*smpi_datatype_size(MPI_CURRENT_TYPE));
 #endif
-   smpi_mpi_reduce(NULL, NULL, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, root, MPI_COMM_WORLD);
+   mpi_coll_reduce_fun(NULL, NULL, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, root, MPI_COMM_WORLD);
    smpi_execute_flops(comp_size);
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__);
@@ -463,19 +460,19 @@ static void action_reduce(const char *const *action)
 static void action_allReduce(const char *const *action) {
   double comm_size = parse_double(action[2]);
   double comp_size = parse_double(action[3]);
-  
+
   if(action[4]) MPI_CURRENT_TYPE=decode_datatype(action[4]);
   else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE;
-  
+
   double clock = smpi_process_simulated_elapsed();
 #ifdef HAVE_TRACING
   int rank = smpi_comm_rank(MPI_COMM_WORLD);
   TRACE_smpi_computing_out(rank);
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,comp_size*smpi_datatype_size(MPI_CURRENT_TYPE));
 #endif
-  smpi_mpi_reduce(NULL, NULL, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, 0, MPI_COMM_WORLD);
+  mpi_coll_reduce_fun(NULL, NULL, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, 0, MPI_COMM_WORLD);
   smpi_execute_flops(comp_size);
-  smpi_mpi_bcast(NULL, comm_size, MPI_CURRENT_TYPE, 0, MPI_COMM_WORLD);
+  mpi_coll_bcast_fun(NULL, comm_size, MPI_CURRENT_TYPE, 0, MPI_COMM_WORLD);
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
   TRACE_smpi_computing_in(rank);
@@ -489,31 +486,79 @@ static void action_allToAll(const char *const *action) {
   int comm_size = smpi_comm_size(MPI_COMM_WORLD);
   int send_size = parse_double(action[2]);
   int recv_size = parse_double(action[3]);
-  void *send = xbt_new0(int, send_size*comm_size);  
-  void *recv = xbt_new0(int, send_size*comm_size);
-  
-  if(action[4]) MPI_CURRENT_TYPE=decode_datatype(action[4]);
-  else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE;
+  MPI_Datatype MPI_CURRENT_TYPE2;
+
+  if(action[4]) {
+    MPI_CURRENT_TYPE=decode_datatype(action[4]);
+    MPI_CURRENT_TYPE2=decode_datatype(action[5]);
+  }
+  else {
+    MPI_CURRENT_TYPE=MPI_DEFAULT_TYPE;
+    MPI_CURRENT_TYPE2=MPI_DEFAULT_TYPE;
+  }
+  void *send = calloc(send_size*comm_size, smpi_datatype_size(MPI_CURRENT_TYPE));  
+  void *recv = calloc(recv_size*comm_size, smpi_datatype_size(MPI_CURRENT_TYPE2));  
 
 #ifdef HAVE_TRACING
   int rank = smpi_process_index();
   TRACE_smpi_computing_out(rank);
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,send_size*smpi_datatype_size(MPI_CURRENT_TYPE));
 #endif
 
-  if (send_size < 200 && comm_size > 12) {
-    smpi_coll_tuned_alltoall_bruck(send, send_size, MPI_CURRENT_TYPE,
-                                   recv, recv_size, MPI_CURRENT_TYPE,
-                                   MPI_COMM_WORLD);
-  } else if (send_size < 3000) {  
-    smpi_coll_tuned_alltoall_basic_linear(send, send_size, MPI_CURRENT_TYPE,
-                                          recv, recv_size, MPI_CURRENT_TYPE,
-                                          MPI_COMM_WORLD);
+  mpi_coll_alltoall_fun(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, MPI_COMM_WORLD);
+
+#ifdef HAVE_TRACING
+  TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
+#endif
+
+  log_timed_action (action, clock);
+  xbt_free(send);
+  xbt_free(recv);
+}
+
+
+static void action_gather(const char *const *action) {
+  /*
+ The structure of the gather action for the rank 0 (total 4 processes) 
+ is the following:   
+ 0 gather 68 68 0 0 0
+
+  where: 
+  1) 68 is the sendcounts
+  2) 68 is the recvcounts
+  3) 0 is the root node
+  4) 0 is the send datatype id, see decode_datatype()
+  5) 0 is the recv datatype id, see decode_datatype()
+  */
+  double clock = smpi_process_simulated_elapsed();
+  int comm_size = smpi_comm_size(MPI_COMM_WORLD);
+  int send_size = parse_double(action[2]);
+  int recv_size = parse_double(action[3]);
+  MPI_Datatype MPI_CURRENT_TYPE2;
+  if(action[5]) {
+    MPI_CURRENT_TYPE=decode_datatype(action[5]);
+    MPI_CURRENT_TYPE2=decode_datatype(action[6]);
   } else {
-    smpi_coll_tuned_alltoall_pairwise(send, send_size, MPI_CURRENT_TYPE,
-                                      recv, recv_size, MPI_CURRENT_TYPE,
-                                      MPI_COMM_WORLD);
+    MPI_CURRENT_TYPE=MPI_DEFAULT_TYPE;
+    MPI_CURRENT_TYPE2=MPI_DEFAULT_TYPE;
   }
+  void *send = calloc(send_size, smpi_datatype_size(MPI_CURRENT_TYPE));
+  void *recv = calloc(recv_size, smpi_datatype_size(MPI_CURRENT_TYPE2));
+
+  int root=atoi(action[4]);
+  int rank = smpi_process_index();
+
+  if(rank==root)
+    recv = calloc(recv_size*comm_size, smpi_datatype_size(MPI_CURRENT_TYPE2));
+
+#ifdef HAVE_TRACING
+  TRACE_smpi_computing_out(rank);
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,send_size*smpi_datatype_size(MPI_CURRENT_TYPE));
+#endif
+smpi_mpi_gather(send, send_size, MPI_CURRENT_TYPE,
+                recv, recv_size, MPI_CURRENT_TYPE2,
+                root, MPI_COMM_WORLD);
 
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
@@ -525,62 +570,185 @@ static void action_allToAll(const char *const *action) {
   xbt_free(recv);
 }
 
+
+static void action_reducescatter(const char *const *action) {
+
+    /*
+ The structure of the reducescatter action for the rank 0 (total 4 processes) 
+ is the following:   
+0 reduceScatter 275427 275427 275427 204020 11346849 0
+
+  where: 
+  1) The first four values after the name of the action declare the recvcounts array
+  2) The value 11346849 is the amount of instructions
+  3) The last value corresponds to the datatype, see decode_datatype().
+
+  We analyze a MPI_Reduce_scatter call to one MPI_Reduce and one MPI_Scatterv.
+
+   */
+
+  double clock = smpi_process_simulated_elapsed();
+  int comm_size = smpi_comm_size(MPI_COMM_WORLD);
+  int comp_size = parse_double(action[2+comm_size]);
+  int *recvcounts = xbt_new0(int, comm_size);  
+  int *disps = xbt_new0(int, comm_size);  
+  int i=0,recv_sum=0;
+  int root=0;
+  int rank = smpi_process_index();
+
+  if(action[3+comm_size])
+    MPI_CURRENT_TYPE=decode_datatype(action[3+comm_size]);
+  else
+    MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE;
+
+  for(i=0;i<comm_size;i++) {
+    recvcounts[i] = atoi(action[i+2]);
+    recv_sum=recv_sum+recvcounts[i];
+    disps[i] = 0;
+  }
+
+#ifdef HAVE_TRACING
+  TRACE_smpi_computing_out(rank);
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, recv_sum*smpi_datatype_size(MPI_CURRENT_TYPE));
+#endif
+   mpi_coll_reduce_fun(NULL, NULL, recv_sum, MPI_CURRENT_TYPE, MPI_OP_NULL,
+       root, MPI_COMM_WORLD);
+   smpi_mpi_scatterv(NULL, recvcounts, disps, MPI_CURRENT_TYPE, NULL,
+                      recvcounts[rank], MPI_CURRENT_TYPE, 0, MPI_COMM_WORLD);
+   smpi_execute_flops(comp_size);
+
+
+#ifdef HAVE_TRACING
+  TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
+#endif
+
+  log_timed_action (action, clock);
+}
+
+
+static void action_allgatherv(const char *const *action) {
+
+  /*
+ The structure of the allgatherv action for the rank 0 (total 4 processes) 
+ is the following:   
+0 allGatherV 275427 275427 275427 275427 204020
+
+  where: 
+  1) 275427 is the sendcount
+  2) The next four elements declare the recvcounts array
+  3) No more values mean that the datatype for sent and receive buffer
+  is the default one, see decode_datatype().
+
+   */
+
+  double clock = smpi_process_simulated_elapsed();
+
+  int comm_size = smpi_comm_size(MPI_COMM_WORLD);
+  int i=0;
+  int sendcount=atoi(action[2]);
+  int *recvcounts = xbt_new0(int, comm_size);  
+  int *disps = xbt_new0(int, comm_size);  
+  int recv_sum=0;  
+  MPI_Datatype MPI_CURRENT_TYPE2;
+
+  if(action[3+comm_size]) {
+    MPI_CURRENT_TYPE = decode_datatype(action[3+comm_size]);
+    MPI_CURRENT_TYPE2 = decode_datatype(action[4+comm_size]);
+  } else {
+    MPI_CURRENT_TYPE = MPI_DEFAULT_TYPE;
+    MPI_CURRENT_TYPE2 = MPI_DEFAULT_TYPE;    
+  }
+  void *sendbuf = calloc(sendcount, smpi_datatype_size(MPI_CURRENT_TYPE));    
+
+  for(i=0;i<comm_size;i++) {
+    recvcounts[i] = atoi(action[i+3]);
+    recv_sum=recv_sum+recvcounts[i];
+  }
+  void *recvbuf = calloc(recv_sum, smpi_datatype_size(MPI_CURRENT_TYPE2));  
+
+#ifdef HAVE_TRACING
+  int rank = MPI_COMM_WORLD != MPI_COMM_NULL ? smpi_process_index() : -1;
+  TRACE_smpi_computing_out(rank);
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,sendcount*smpi_datatype_size(MPI_CURRENT_TYPE));
+#endif
+
+mpi_coll_allgatherv_fun(sendbuf, sendcount, MPI_CURRENT_TYPE, recvbuf, recvcounts, disps, MPI_CURRENT_TYPE2, MPI_COMM_WORLD);
+
+#ifdef HAVE_TRACING
+  TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
+  TRACE_smpi_computing_in(rank);
+#endif
+
+  log_timed_action (action, clock);
+  xbt_free(sendbuf);
+  xbt_free(recvbuf);
+  xbt_free(recvcounts);
+  xbt_free(disps);
+}
+
+
 static void action_allToAllv(const char *const *action) {
   /*
  The structure of the allToAllV action for the rank 0 (total 4 processes) 
  is the following:   
-  0 allToAllV 100 1 7 10 12 5 10 20 45 100 1 70 10 5 1 5 77 90
+  0 allToAllV 100 1 7 10 12 100 1 70 10 5
 
   where: 
   1) 100 is the size of the send buffer *sizeof(int),
   2) 1 7 10 12 is the sendcounts array
-  3) 5 10 20 45 is the sdispls array
-  4) 100*sizeof(int) is the size of the receiver buffer
-  5)  1 70 10 5 is the recvcounts array
-  6) 1 5 77 90 is the rdispls array
-    
+  3) 100*sizeof(int) is the size of the receiver buffer
+  4)  1 70 10 5 is the recvcounts array
+
    */
-  
-  
+
+
   double clock = smpi_process_simulated_elapsed();
-  
+
   int comm_size = smpi_comm_size(MPI_COMM_WORLD);
   int send_buf_size=0,recv_buf_size=0,i=0;
   int *sendcounts = xbt_new0(int, comm_size);  
   int *recvcounts = xbt_new0(int, comm_size);  
   int *senddisps = xbt_new0(int, comm_size);  
   int *recvdisps = xbt_new0(int, comm_size);  
-  
-  send_buf_size=parse_double(action[2]);
-  recv_buf_size=parse_double(action[3+2*comm_size]);
 
-  int *sendbuf = xbt_new0(int, send_buf_size);  
-  int *recvbuf = xbt_new0(int, recv_buf_size);  
+  MPI_Datatype MPI_CURRENT_TYPE2;
 
-  if(action[4+4*comm_size]) MPI_CURRENT_TYPE=decode_datatype(action[4+4*comm_size]);    
-  else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE;
+  send_buf_size=parse_double(action[2]);
+  recv_buf_size=parse_double(action[3+comm_size]);
+  if(action[4+2*comm_size]) {
+    MPI_CURRENT_TYPE=decode_datatype(action[4+2*comm_size]);
+    MPI_CURRENT_TYPE2=decode_datatype(action[5+2*comm_size]);
+  }
+  else {
+      MPI_CURRENT_TYPE=MPI_DEFAULT_TYPE;
+      MPI_CURRENT_TYPE2=MPI_DEFAULT_TYPE;
+  }
+
+  void *sendbuf = calloc(send_buf_size, smpi_datatype_size(MPI_CURRENT_TYPE));  
+  void *recvbuf = calloc(recv_buf_size, smpi_datatype_size(MPI_CURRENT_TYPE2));  
 
   for(i=0;i<comm_size;i++) {
     sendcounts[i] = atoi(action[i+3]);
-    senddisps[i] = atoi(action[i+3+comm_size]);
-    recvcounts[i] = atoi(action[i+4+2*comm_size]);
-    recvdisps[i] = atoi(action[i+4+3*comm_size]);
+    recvcounts[i] = atoi(action[i+4+comm_size]);
   }
-  
+
 
 #ifdef HAVE_TRACING
   int rank = MPI_COMM_WORLD != MPI_COMM_NULL ? smpi_process_index() : -1;
   TRACE_smpi_computing_out(rank);
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
+  int count=0;
+  for(i=0;i<comm_size;i++) count+=sendcounts[i];
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,count*smpi_datatype_size(MPI_CURRENT_TYPE));
 #endif
-    smpi_coll_basic_alltoallv(sendbuf, sendcounts, senddisps, 	MPI_CURRENT_TYPE,
+    mpi_coll_alltoallv_fun(sendbuf, sendcounts, senddisps, MPI_CURRENT_TYPE,
                                recvbuf, recvcounts, recvdisps, MPI_CURRENT_TYPE,
                                MPI_COMM_WORLD);
 #ifdef HAVE_TRACING
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
   TRACE_smpi_computing_in(rank);
 #endif
-   
+
   log_timed_action (action, clock);
   xbt_free(sendbuf);
   xbt_free(recvbuf);
@@ -588,8 +756,6 @@ static void action_allToAllv(const char *const *action) {
   xbt_free(recvcounts);
   xbt_free(senddisps);
   xbt_free(recvdisps);
-
-  
 }
 
 void smpi_replay_init(int *argc, char***argv){
@@ -613,6 +779,9 @@ void smpi_replay_init(int *argc, char***argv){
     xbt_replay_action_register("allReduce",  action_allReduce);
     xbt_replay_action_register("allToAll",   action_allToAll);
     xbt_replay_action_register("allToAllV",  action_allToAllv);
+    xbt_replay_action_register("gather",  action_gather);
+    xbt_replay_action_register("allGatherV",  action_allgatherv);
+    xbt_replay_action_register("reduceScatter",  action_reducescatter);
     xbt_replay_action_register("compute",    action_compute);
   }
 
@@ -623,13 +792,18 @@ int smpi_replay_finalize(){
   double sim_time= 1.;
   /* One active process will stop. Decrease the counter*/
   active_processes--;
+  XBT_DEBUG("There are %lu elements in reqq[*]",
+            xbt_dynar_length(reqq[smpi_comm_rank(MPI_COMM_WORLD)]));
+  xbt_dynar_free(&reqq[smpi_comm_rank(MPI_COMM_WORLD)]);
   if(!active_processes){
     /* Last process alive speaking */
     /* end the simulated timer */
-    xbt_dynar_free(reqq);
     sim_time = smpi_process_simulated_elapsed();
     XBT_INFO("Simulation time %g", sim_time);
     _xbt_replay_action_exit();
+    xbt_free(reqq);
+    reqq = NULL;
   }
+  smpi_mpi_barrier(MPI_COMM_WORLD);
   return PMPI_Finalize();
 }