Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
* added support for optimized collectives:
[simgrid.git] / src / smpi / smpi_mpi.c
index b29243e..9267288 100644 (file)
-// FIXME: remove
-#include <stdio.h>
+
+
 #include "private.h"
+#include "smpi_coll_private.h"
+
+XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi, smpi,
+                                "Logging specific to SMPI (mpi)");
 
-int MPI_Init(int *argc, char ***argv)
+int SMPI_MPI_Init(int *argc, char ***argv)
 {
-       smpi_mpi_init();
-       smpi_bench_begin();
-       return MPI_SUCCESS;
+  smpi_process_init(argc, argv);
+  smpi_bench_begin();
+  return MPI_SUCCESS;
 }
 
-int MPI_Finalize()
+int SMPI_MPI_Finalize()
 {
-       smpi_bench_end();
-       smpi_mpi_finalize();
-       return MPI_SUCCESS;
+  smpi_bench_end();
+  smpi_process_finalize();
+  return MPI_SUCCESS;
 }
 
 // right now this just exits the current node, should send abort signal to all
-// hosts in the communicator;
-int MPI_Abort(MPI_Comm comm, int errorcode)
+// hosts in the communicator (TODO)
+int SMPI_MPI_Abort(MPI_Comm comm, int errorcode)
 {
-       smpi_exit(errorcode);
-       return 0;
+  smpi_exit(errorcode);
+  return 0;
 }
 
-int MPI_Comm_size(MPI_Comm comm, int *size)
+int SMPI_MPI_Comm_size(MPI_Comm comm, int *size)
 {
-       int retval = MPI_SUCCESS;
+  int retval = MPI_SUCCESS;
 
-       smpi_bench_end();
+  smpi_bench_end();
 
-       if (NULL == comm) {
-               retval = MPI_ERR_COMM;
-       } else if (NULL == size) {
-               retval = MPI_ERR_ARG;
-       } else {
-               *size = comm->size;
-       }
+  if (NULL == comm) {
+    retval = MPI_ERR_COMM;
+  } else if (NULL == size) {
+    retval = MPI_ERR_ARG;
+  } else {
+    *size = comm->size;
+  }
 
-       smpi_bench_begin();
+  smpi_bench_begin();
 
-       return retval;
+  return retval;
 }
 
-int MPI_Comm_rank(MPI_Comm comm, int *rank)
+int SMPI_MPI_Comm_rank(MPI_Comm comm, int *rank)
 {
-       int retval = MPI_SUCCESS;
+  int retval = MPI_SUCCESS;
 
-       smpi_bench_end();
+  smpi_bench_end();
 
-       if (NULL == comm) {
-               retval = MPI_ERR_COMM;
-       } else if (NULL == rank) {
-               retval = MPI_ERR_ARG;
-       } else {
-               *rank = smpi_mpi_comm_rank(comm);
-       }
+  if (NULL == comm) {
+    retval = MPI_ERR_COMM;
+  } else if (NULL == rank) {
+    retval = MPI_ERR_ARG;
+  } else {
+    *rank = smpi_mpi_comm_rank(comm);
+  }
 
-       smpi_bench_begin();
+  smpi_bench_begin();
 
-       return retval;
+  return retval;
 }
 
-int MPI_Type_size(MPI_Datatype datatype, size_t *size)
+int SMPI_MPI_Type_size(MPI_Datatype datatype, size_t * size)
 {
-       int retval = MPI_SUCCESS;
+  int retval = MPI_SUCCESS;
 
-       smpi_bench_end();
+  smpi_bench_end();
 
-       if (NULL == datatype) {
-               retval = MPI_ERR_TYPE;
-       } else if (NULL == size) {
-               retval = MPI_ERR_ARG;
-       } else {
-               *size = datatype->size;
-       }
+  if (NULL == datatype) {
+    retval = MPI_ERR_TYPE;
+  } else if (NULL == size) {
+    retval = MPI_ERR_ARG;
+  } else {
+    *size = datatype->size;
+  }
 
-       smpi_bench_begin();
+  smpi_bench_begin();
 
-       return retval;
+  return retval;
 }
 
-int MPI_Barrier(MPI_Comm comm)
+int SMPI_MPI_Barrier(MPI_Comm comm)
 {
-       int retval = MPI_SUCCESS;
+  int retval = MPI_SUCCESS;
 
-       smpi_bench_end();
+  smpi_bench_end();
 
-       if (NULL == comm) {
-               retval = MPI_ERR_COMM;
-       } else {
-               retval = smpi_mpi_barrier(comm);
-       }
+  if (NULL == comm) {
+    retval = MPI_ERR_COMM;
+  } else {
+    retval = smpi_mpi_barrier(comm);
+  }
 
-       smpi_bench_begin();
+  smpi_bench_begin();
 
-       return retval;
+  return retval;
 }
 
-int MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Request *request)
+int SMPI_MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src,
+                   int tag, MPI_Comm comm, MPI_Request * request)
 {
-       int retval = MPI_SUCCESS;
+  int retval = MPI_SUCCESS;
 
-       smpi_bench_end();
+  smpi_bench_end();
 
-       retval = smpi_create_request(buf, count, datatype, src, 0, tag, comm, request);
-       if (NULL != *request && MPI_SUCCESS == retval) {
-               retval = smpi_mpi_irecv(*request);
-       }
+  retval = smpi_create_request(buf, count, datatype, src, 0, tag, comm,
+                               request);
+  if (NULL != *request && MPI_SUCCESS == retval) {
+    retval = smpi_mpi_irecv(*request);
+  }
 
-       smpi_bench_begin();
+  smpi_bench_begin();
 
-       return retval;
+  return retval;
 }
 
-int MPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status *status)
+int SMPI_MPI_Recv(void *buf, int count, MPI_Datatype datatype, int src,
+                  int tag, MPI_Comm comm, MPI_Status * status)
 {
-       int retval = MPI_SUCCESS;
-       smpi_mpi_request_t request;
+  int retval = MPI_SUCCESS;
+  smpi_mpi_request_t request;
 
-       smpi_bench_end();
+  smpi_bench_end();
 
-       retval = smpi_create_request(buf, count, datatype, src, 0, tag, comm, &request);
-       if (NULL != request && MPI_SUCCESS == retval) {
-               retval = smpi_mpi_irecv(request);
-               if (MPI_SUCCESS == retval) {
-                       retval = smpi_mpi_wait(request, status);
-               }
-               xbt_mallocator_release(smpi_global->request_mallocator, request);
-       }
+  retval = smpi_create_request(buf, count, datatype, src, 0, tag, comm,
+                               &request);
+  if (NULL != request && MPI_SUCCESS == retval) {
+    retval = smpi_mpi_irecv(request);
+    if (MPI_SUCCESS == retval) {
+      retval = smpi_mpi_wait(request, status);
+    }
+    xbt_mallocator_release(smpi_global->request_mallocator, request);
+  }
 
-       smpi_bench_begin();
+  smpi_bench_begin();
 
-       return retval;
+  return retval;
 }
 
-int MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm, MPI_Request *request)
+int SMPI_MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst,
+                   int tag, MPI_Comm comm, MPI_Request * request)
 {
-       int retval = MPI_SUCCESS;
+  int retval = MPI_SUCCESS;
+
+  smpi_bench_end();
+
+  retval = smpi_create_request(buf, count, datatype, 0, dst, tag, comm,
+                               request);
+  if (NULL != *request && MPI_SUCCESS == retval) {
+    retval = smpi_mpi_isend(*request);
+  }
+
+  smpi_bench_begin();
+
+  return retval;
+}
+
+int SMPI_MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst,
+                  int tag, MPI_Comm comm)
+{
+  int retval = MPI_SUCCESS;
+  smpi_mpi_request_t request;
+
+  smpi_bench_end();
 
-       smpi_bench_end();
+  retval = smpi_create_request(buf, count, datatype, 0, dst, tag, comm,
+                               &request);
+  if (NULL != request && MPI_SUCCESS == retval) {
+    retval = smpi_mpi_isend(request);
+    if (MPI_SUCCESS == retval) {
+      smpi_mpi_wait(request, MPI_STATUS_IGNORE);
+    }
+    xbt_mallocator_release(smpi_global->request_mallocator, request);
+  }
 
-       retval = smpi_create_request(buf, count, datatype, 0, dst, tag, comm, request);
-       if (NULL != *request && MPI_SUCCESS == retval) {
-               retval = smpi_mpi_isend(*request);
-       }
+  smpi_bench_begin();
 
-       smpi_bench_begin();
+  return retval;
+}
 
-       return retval;
+/**
+ * MPI_Wait and friends
+ **/
+int SMPI_MPI_Wait(MPI_Request * request, MPI_Status * status)
+{
+  return smpi_mpi_wait(*request, status);
 }
 
-int MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+int SMPI_MPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
 {
-       int retval = MPI_SUCCESS;
-       smpi_mpi_request_t request;
+  return smpi_mpi_waitall(count, requests, status);
+}
 
-       smpi_bench_end();
+int SMPI_MPI_Waitany(int count, MPI_Request requests[], int *index,
+                     MPI_Status status[])
+{
+  return smpi_mpi_waitany(count, requests, index, status);
+}
 
-       retval = smpi_create_request(buf, count, datatype, 0, dst, tag, comm, &request);
-       if (NULL != request && MPI_SUCCESS == retval) {
-               retval = smpi_mpi_isend(request);
-               if (MPI_SUCCESS == retval) {
-                       smpi_mpi_wait(request, MPI_STATUS_IGNORE);
-               }
-               xbt_mallocator_release(smpi_global->request_mallocator, request);
-       }
+/**
+ * MPI_Bcast
+ **/
 
-       smpi_bench_begin();
+/**
+ * flat bcast 
+ **/
+int flat_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
+int flat_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root,
+                MPI_Comm comm)
+{
+        int rank;
+        int retval = MPI_SUCCESS;
+        smpi_mpi_request_t request;
+
+        rank = smpi_mpi_comm_rank(comm);
+        if (rank == root) {
+                retval = smpi_create_request(buf, count, datatype, root,
+                                (root + 1) % comm->size, 0, comm, &request);
+                request->forward = comm->size - 1;
+                smpi_mpi_isend(request);
+        } else {
+                retval = smpi_create_request(buf, count, datatype, MPI_ANY_SOURCE, rank,
+                                0, comm, &request);
+                smpi_mpi_irecv(request);
+        }
+
+        smpi_mpi_wait(request, MPI_STATUS_IGNORE);
+        xbt_mallocator_release(smpi_global->request_mallocator, request);
+
+        return(retval);
 
-       return retval;
 }
 
-int MPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm) {
+/**
+ * Bcast user entry point
+ **/
+int SMPI_MPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root,
+                   MPI_Comm comm)
+{
+
+  int retval = MPI_SUCCESS;
 
-       int retval = MPI_SUCCESS;
-       int rank;
-       smpi_mpi_request_t request;
+  smpi_bench_end();
 
-       smpi_bench_end();
+  //retval = flat_tree_bcast(buf, count, datatype, root, comm);
+  retval = binomial_tree_bcast(buf, count, datatype, root, comm);
 
-       rank = smpi_mpi_comm_rank(comm);
+  smpi_bench_begin();
 
-       if (rank == root) {
-               retval = smpi_create_request(buf, count, datatype, root, (root + 1) % comm->size, 0, comm, &request);
-               request->forward = comm->size - 1;
-               smpi_mpi_isend(request);
-       } else {
-               retval = smpi_create_request(buf, count, datatype, MPI_ANY_SOURCE, rank, 0, comm, &request);
-               smpi_mpi_irecv(request);
-       }
+  return retval;
+}
 
-       smpi_mpi_wait(request, MPI_STATUS_IGNORE);
-       xbt_mallocator_release(smpi_global->request_mallocator, request);
 
-       smpi_bench_begin();
 
-       return retval;
+#ifdef DEBUG_REDUCE
+/**
+ * debugging helper function
+ **/
+static void print_buffer_int(void *buf, int len, const char *msg, int rank)
+{
+  int tmp, *v;
+  printf("**[%d] %s: ", rank, msg);
+  for (tmp = 0; tmp < len; tmp++) {
+    v = buf;
+    printf("[%d]", v[tmp]);
+  }
+  printf("\n");
+  free(msg);
+}
+#endif
+
+/**
+ * MPI_Reduce
+ **/
+int SMPI_MPI_Reduce(void *sendbuf, void *recvbuf, int count,
+                    MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm)
+{
+  int retval = MPI_SUCCESS;
+  int rank;
+  int size;
+  int i;
+  int tag = 0;
+  smpi_mpi_request_t *tabrequest;
+  smpi_mpi_request_t request;
+
+  smpi_bench_end();
+
+  rank = smpi_mpi_comm_rank(comm);
+  size = comm->size;
+
+  if (rank != root) {           // if i am not ROOT, simply send my buffer to root
+
+#ifdef DEBUG_REDUCE
+    print_buffer_int(sendbuf, count, xbt_strdup("sndbuf"), rank);
+#endif
+    retval =
+      smpi_create_request(sendbuf, count, datatype, rank, root, tag, comm,
+                          &request);
+    smpi_mpi_isend(request);
+    smpi_mpi_wait(request, MPI_STATUS_IGNORE);
+    xbt_mallocator_release(smpi_global->request_mallocator, request);
+
+  } else {
+    // i am the ROOT: wait for all buffers by creating one request by sender
+    int src;
+    tabrequest = xbt_malloc((size - 1) * sizeof(smpi_mpi_request_t));
+
+    void **tmpbufs = xbt_malloc((size - 1) * sizeof(void *));
+    for (i = 0; i < size - 1; i++) {
+      // we need 1 buffer per request to store intermediate receptions
+      tmpbufs[i] = xbt_malloc(count * datatype->size);
+    }
+    memcpy(recvbuf, sendbuf, count * datatype->size * sizeof(char));    // initiliaze recv buf with my own snd buf
+
+    // i can not use: 'request->forward = size-1;' (which would progagate size-1 receive reqs)
+    // since we should op values as soon as one receiving request matches.
+    for (i = 0; i < size - 1; i++) {
+      // reminder: for smpi_create_request() the src is always the process sending.
+      src = i < root ? i : i + 1;
+      retval = smpi_create_request(tmpbufs[i], count, datatype,
+                                   src, root, tag, comm, &(tabrequest[i]));
+      if (NULL != tabrequest[i] && MPI_SUCCESS == retval) {
+        if (MPI_SUCCESS == retval) {
+          smpi_mpi_irecv(tabrequest[i]);
+        }
+      }
+    }
+    // now, wait for completion of all irecv's.
+    for (i = 0; i < size - 1; i++) {
+      int index = MPI_UNDEFINED;
+      smpi_mpi_waitany(size - 1, tabrequest, &index, MPI_STATUS_IGNORE);
+
+#ifdef DEBUG_REDUCE
+      printf
+        ("MPI_Waitany() unblocked: root received (completes req[index=%d])\n",
+         index);
+      print_buffer_int(tmpbufs[index], count,
+                       bprintf("tmpbufs[index=%d] (value received)", index),
+                       rank);
+#endif
+
+      // arg 2 is modified
+      op->func(tmpbufs[index], recvbuf, &count, &datatype);
+#ifdef DEBUG_REDUCE
+      print_buffer_int(recvbuf, count, xbt_strdup("rcvbuf"), rank);
+
+#endif
+      //xbt_mallocator_release(smpi_global->request_mallocator, tabrequest[i]);
+      xbt_free(tmpbufs[index]);
+    }
+    xbt_free(tabrequest);
+    xbt_free(tmpbufs);
+  }
+
+  smpi_bench_begin();
+
+  return retval;
+}
+
+/**
+ * MPI_Allreduce
+ *
+ * Same as MPI_REDUCE except that the result appears in the receive buffer of all the group members.
+ **/
+int SMPI_MPI_Allreduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
+                        MPI_Op op, MPI_Comm comm );
+int SMPI_MPI_Allreduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
+                        MPI_Op op, MPI_Comm comm )
+{
+  int retval = MPI_SUCCESS;
+  int root=0;  // arbitrary choice
+
+  smpi_bench_end();
+
+  retval = SMPI_MPI_Reduce( sendbuf, recvbuf, count, datatype, op, root, comm);
+  if (MPI_SUCCESS != retval)
+           return(retval);
+  retval = SMPI_MPI_Bcast( recvbuf, count, datatype, root, comm);
+
+  smpi_bench_begin();
+  return( retval );
+}
+
+
+
+
+
+// used by comm_split to sort ranks based on key values
+int smpi_compare_rankkeys(const void *a, const void *b);
+int smpi_compare_rankkeys(const void *a, const void *b)
+{
+  int *x = (int *) a;
+  int *y = (int *) b;
+
+  if (x[1] < y[1])
+    return -1;
+
+  if (x[1] == y[1]) {
+    if (x[0] < y[0])
+      return -1;
+    if (x[0] == y[0])
+      return 0;
+    return 1;
+  }
+
+  return 1;
+}
+
+int SMPI_MPI_Comm_split(MPI_Comm comm, int color, int key,
+                        MPI_Comm * comm_out)
+{
+  int retval = MPI_SUCCESS;
+
+  int index, rank;
+  smpi_mpi_request_t request;
+  int colorkey[2];
+  smpi_mpi_status_t status;
+
+  smpi_bench_end();
+
+  // FIXME: need to test parameters
+
+  index = smpi_process_index();
+  rank = comm->index_to_rank_map[index];
+
+  // default output
+  comm_out = NULL;
+
+  // root node does most of the real work
+  if (0 == rank) {
+    int colormap[comm->size];
+    int keymap[comm->size];
+    int rankkeymap[comm->size * 2];
+    int i, j;
+    smpi_mpi_communicator_t tempcomm = NULL;
+    int count;
+    int indextmp;
+
+    colormap[0] = color;
+    keymap[0] = key;
+
+    // FIXME: use scatter/gather or similar instead of individual comms
+    for (i = 1; i < comm->size; i++) {
+      retval = smpi_create_request(colorkey, 2, MPI_INT, MPI_ANY_SOURCE,
+                                   rank, MPI_ANY_TAG, comm, &request);
+      smpi_mpi_irecv(request);
+      smpi_mpi_wait(request, &status);
+      colormap[status.MPI_SOURCE] = colorkey[0];
+      keymap[status.MPI_SOURCE] = colorkey[1];
+      xbt_mallocator_release(smpi_global->request_mallocator, request);
+    }
+
+    for (i = 0; i < comm->size; i++) {
+      if (MPI_UNDEFINED == colormap[i]) {
+        continue;
+      }
+      // make a list of nodes with current color and sort by keys
+      count = 0;
+      for (j = i; j < comm->size; j++) {
+        if (colormap[i] == colormap[j]) {
+          colormap[j] = MPI_UNDEFINED;
+          rankkeymap[count * 2] = j;
+          rankkeymap[count * 2 + 1] = keymap[j];
+          count++;
+        }
+      }
+      qsort(rankkeymap, count, sizeof(int) * 2, &smpi_compare_rankkeys);
+
+      // new communicator
+      tempcomm = xbt_new(s_smpi_mpi_communicator_t, 1);
+      tempcomm->barrier_count = 0;
+      tempcomm->size = count;
+      tempcomm->barrier_mutex = SIMIX_mutex_init();
+      tempcomm->barrier_cond = SIMIX_cond_init();
+      tempcomm->rank_to_index_map = xbt_new(int, count);
+      tempcomm->index_to_rank_map = xbt_new(int, smpi_global->process_count);
+      for (j = 0; j < smpi_global->process_count; j++) {
+        tempcomm->index_to_rank_map[j] = -1;
+      }
+      for (j = 0; j < count; j++) {
+        indextmp = comm->rank_to_index_map[rankkeymap[j * 2]];
+        tempcomm->rank_to_index_map[j] = indextmp;
+        tempcomm->index_to_rank_map[indextmp] = j;
+      }
+      for (j = 0; j < count; j++) {
+        if (rankkeymap[j * 2]) {
+          retval = smpi_create_request(&j, 1, MPI_INT, 0,
+                                       rankkeymap[j * 2], 0, comm, &request);
+          request->data = tempcomm;
+          smpi_mpi_isend(request);
+          smpi_mpi_wait(request, &status);
+          xbt_mallocator_release(smpi_global->request_mallocator, request);
+        } else {
+          *comm_out = tempcomm;
+        }
+      }
+    }
+  } else {
+    colorkey[0] = color;
+    colorkey[1] = key;
+    retval = smpi_create_request(colorkey, 2, MPI_INT, rank, 0, 0, comm,
+                                 &request);
+    smpi_mpi_isend(request);
+    smpi_mpi_wait(request, &status);
+    xbt_mallocator_release(smpi_global->request_mallocator, request);
+    if (MPI_UNDEFINED != color) {
+      retval = smpi_create_request(colorkey, 1, MPI_INT, 0, rank, 0, comm,
+                                   &request);
+      smpi_mpi_irecv(request);
+      smpi_mpi_wait(request, &status);
+      *comm_out = request->data;
+    }
+  }
+
+  smpi_bench_begin();
+
+  return retval;
 }
 
-// FIXME: needs to return null in event of MPI_UNDEFINED color...
-// FIXME: seriously, this isn't pretty
-int MPI_Comm_split(MPI_Comm comm, int color, int key, MPI_Comm *comm_out)
+double SMPI_MPI_Wtime(void)
 {
-       int retval = MPI_SUCCESS;
-
-       int index, rank;
-       smpi_mpi_request_t request;
-       int colorkey[2];
-       smpi_mpi_status_t status;
-
-       smpi_bench_end();
-
-       // FIXME: need to test parameters
-
-       index = smpi_host_index();
-       rank  = comm->index_to_rank_map[index];
-
-       if (0 == rank) {
-               int *colors = xbt_new(int, comm->size);
-               int *keys   = xbt_new(int, comm->size);
-               int i, j, k;
-               smpi_mpi_communicator_t tempcomm = NULL;
-               int colortmp;
-               int keycount;
-               int *keystmp  = xbt_new(int, comm->size);
-               int *rankstmp = xbt_new(int, comm->size);
-               int tmpval;
-               int indextmp;
-
-               colors[0] = color;
-               keys[0]   = key;
-
-               // FIXME: not efficient
-               for (i = 1; i < comm->size; i++) {
-                       retval = smpi_create_request(colorkey, 2, MPI_INT, MPI_ANY_SOURCE, rank, MPI_ANY_TAG, comm, &request);
-                       smpi_mpi_irecv(request);
-                       smpi_mpi_wait(request, &status);
-                       xbt_mallocator_release(smpi_global->request_mallocator, request);
-                       colors[i] = colorkey[0];
-                       keys[i]   = colorkey[1];
-               }
-
-               for (i = 0; i < comm->size; i++) {
-                       if (-1 == colors[i]) {
-                               continue;
-                       }
-                       colortmp = colors[i];
-                       keycount = 0;
-                       for (j = i; j < comm->size; j++) {
-                               if(colortmp == colors[j]) {
-                                       colors[j] = -1;
-                                       keystmp[keycount] = keys[j];
-                                       rankstmp[keycount] = j;
-                                       keycount++;
-                               }
-                       }
-                       if (0 < keycount) {
-                               // FIXME: yes, mock me, bubble sort...
-                               for (j = 0; j < keycount; j++) {
-                                       for (k = keycount - 1; k > j; k--) {
-                                               if (keystmp[k] < keystmp[k - 1]) {
-                                                       tmpval          = keystmp[k];
-                                                       keystmp[k]      = keystmp[k - 1];
-                                                       keystmp[k - 1]  = tmpval;
-
-                                                       tmpval          = rankstmp[k];
-                                                       rankstmp[k]     = rankstmp[k - 1];
-                                                       rankstmp[k - 1] = tmpval;
-                                               }
-                                       }
-                               }
-                               tempcomm                    = xbt_new(s_smpi_mpi_communicator_t, 1);
-                               tempcomm->barrier_count     = 0;
-                               tempcomm->barrier_mutex     = SIMIX_mutex_init();
-                               tempcomm->barrier_cond      = SIMIX_cond_init();
-                               tempcomm->rank_to_index_map = xbt_new(int, keycount);
-                               tempcomm->index_to_rank_map = xbt_new(int, smpi_global->host_count);
-                               for (j = 0; j < smpi_global->host_count; j++) {
-                                       tempcomm->index_to_rank_map[j] = -1;
-                               }
-                               for (j = 0; j < keycount; j++) {
-                                       indextmp = comm->rank_to_index_map[rankstmp[j]];
-                                       tempcomm->rank_to_index_map[j]        = indextmp;
-                                       tempcomm->index_to_rank_map[indextmp] = j;
-                               }
-                               for (j = 0; j < keycount; j++) {
-                                       retval = smpi_create_request(&j, 1, MPI_INT, 0, rankstmp[j], 0, comm, &request);
-                                       request->data = tempcomm;
-                                       smpi_mpi_isend(request);
-                                       smpi_mpi_wait(request, &status);
-                                       xbt_mallocator_release(smpi_global->request_mallocator, request);
-                               }
-                       }
-               }
-       } else {
-               colorkey[0] = color;
-               colorkey[1] = key;
-               retval = smpi_create_request(colorkey, 2, MPI_INT, rank, 0, 0, comm, &request);
-               smpi_mpi_isend(request);
-               smpi_mpi_wait(request, &status);
-               xbt_mallocator_release(smpi_global->request_mallocator, request);
-               retval = smpi_create_request(colorkey, 1, MPI_INT, 0, rank, 0, comm, &request);
-               smpi_mpi_irecv(request);
-               smpi_mpi_wait(request, &status);
-               *comm_out = request->data;
-       }
-
-       smpi_bench_begin();
-
-       return retval;
+  return (SIMIX_get_clock());
 }