X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/16e1615dff6a9f5a19efb4861563436a638642ed..6efc4db71d869ae5af39d17eeeff2a615aba3ccf:/src/smpi/smpi_base.c

diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c
index fbc34fb0a2..2e192a05ae 100644
--- a/src/smpi/smpi_base.c
+++ b/src/smpi/smpi_base.c
@@ -1,3 +1,9 @@
+/* Copyright (c) 2007, 2008, 2009, 2010. 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. */
+
 #include "private.h"
 #include "xbt/time.h"
 
@@ -7,194 +13,631 @@ XBT_LOG_EXTERNAL_CATEGORY(smpi_base);
 XBT_LOG_EXTERNAL_CATEGORY(smpi_bench);
 XBT_LOG_EXTERNAL_CATEGORY(smpi_kernel);
 XBT_LOG_EXTERNAL_CATEGORY(smpi_mpi);
+XBT_LOG_EXTERNAL_CATEGORY(smpi_mpi_dt);
+XBT_LOG_EXTERNAL_CATEGORY(smpi_coll);
 XBT_LOG_EXTERNAL_CATEGORY(smpi_receiver);
 XBT_LOG_EXTERNAL_CATEGORY(smpi_sender);
 XBT_LOG_EXTERNAL_CATEGORY(smpi_util);
 
-smpi_mpi_global_t smpi_mpi_global = NULL;
+#define EAGER_LIMIT 65536
+#define RDV_TAG     (-10)
+
+void smpi_process_init(int* argc, char*** argv) {
+  int index;
+  smpi_process_data_t data;
+  smx_process_t proc;
+
+  proc = SIMIX_process_self();
+  index = atoi((*argv)[1]);
+  data = smpi_process_remote_data(index);
+  SIMIX_process_set_data(proc, data);
+  if (*argc > 2) {
+    free((*argv)[1]);
+    memmove(&(*argv)[1], &(*argv)[2], sizeof(char *) * (*argc - 2));
+    (*argv)[(*argc) - 1] = NULL;
+  }
+  (*argc)--;
+  DEBUG2("<%d> New process in the game: %p", index, proc);
+}
 
-void smpi_mpi_land_func(void *a, void *b, int *length,
-                        MPI_Datatype * datatype);
+void smpi_process_destroy(void) {
+  int index = smpi_process_index();
 
-void smpi_mpi_land_func(void *a, void *b, int *length,
-                        MPI_Datatype * datatype)
-{
-  int i;
-  if (*datatype == smpi_mpi_global->mpi_int) {
-    int *x = a, *y = b;
-    for (i = 0; i < *length; i++) {
-      y[i] = x[i] && y[i];
-    }
+  DEBUG1("<%d> Process left the game", index);
+}
+
+static MPI_Request build_request(void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags) {
+  MPI_Request request;
+
+  request = xbt_new(s_smpi_mpi_request_t, 1);
+  request->buf = buf;
+  request->size = smpi_datatype_size(datatype) * count;
+  request->src = src;
+  request->dst = dst;
+  request->tag = tag;
+  request->comm = comm;
+  request->rdv = NULL;
+  request->pair = NULL;
+  request->complete = 0;
+  request->match = MPI_REQUEST_NULL;
+  request->flags = flags;
+  if(request->size < EAGER_LIMIT) {
+    request->ack = MPI_REQUEST_NULL;
+  } else {
+    request->ack = xbt_new(s_smpi_mpi_request_t, 1);
+    request->ack->buf = NULL;
+    request->ack->size = 0;
+    request->ack->src = dst;
+    request->ack->dst = src;
+    request->ack->tag = RDV_TAG;
+    request->ack->comm = comm;
+    request->ack->rdv = NULL;
+    request->ack->pair = NULL;
+    request->ack->complete = 0;
+    request->ack->match = MPI_REQUEST_NULL;
+    request->ack->flags = NON_PERSISTENT | ((request->flags & RECV) == RECV ? SEND : RECV);
+    smpi_mpi_start(request->ack);
   }
+  return request;
+}
+
+/* MPI Low level calls */
+MPI_Request smpi_mpi_send_init(void* buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) {
+  MPI_Request request = build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, comm, PERSISTENT | SEND);
+
+  return request;
+}
+
+MPI_Request smpi_mpi_recv_init(void* buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) {
+  MPI_Request request = build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag, comm, PERSISTENT | RECV);
+
+  return request;
 }
 
-void smpi_mpi_sum_func(void *a, void *b, int *length,
-                       MPI_Datatype * datatype);
+void smpi_mpi_start(MPI_Request request) {
+  xbt_assert0(request->complete == 0, "Cannot start a non-finished communication");
+  if(request->size >= EAGER_LIMIT) {
+    print_request("RDV ack", request->ack);
+    smpi_mpi_wait(&request->ack, MPI_STATUS_IGNORE);
+  }
+  if((request->flags & RECV) == RECV) {
+    smpi_process_post_recv(request);
+    print_request("New recv", request);
+    request->pair = SIMIX_network_irecv(request->rdv, request->buf, &request->size);
+  } else {
+    smpi_process_post_send(request->comm, request); // FIXME
+    print_request("New send", request);
+    request->pair = SIMIX_network_isend(request->rdv, request->size, -1.0, request->buf, request->size, NULL);
+  }
+}
 
-void smpi_mpi_sum_func(void *a, void *b, int *length, MPI_Datatype * datatype)
-{
+void smpi_mpi_startall(int count, MPI_Request* requests) {
   int i;
-  if (*datatype == smpi_mpi_global->mpi_int) {
-    int *x = a, *y = b;
-    for (i = 0; i < *length; i++) {
-      y[i] = x[i] + y[i];
-    }
+
+  for(i = 0; i < count; i++) {
+    smpi_mpi_start(requests[i]);
   }
 }
 
-int smpi_mpi_comm_rank(smpi_mpi_communicator_t comm)
-{
-  return comm->index_to_rank_map[smpi_host_index()];
+void smpi_mpi_request_free(MPI_Request* request) {
+  xbt_free(*request);
+  *request = MPI_REQUEST_NULL;
 }
 
-void smpi_process_init()
-{
-  smx_host_t host;
-  int i;
-  smpi_host_data_t hdata;
+MPI_Request smpi_isend_init(void* buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) {
+  MPI_Request request = build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, comm, NON_PERSISTENT | SEND);
 
-  smpi_global->running_hosts_count++;
+  return request;
+}
 
-  // initialize some local variables
-  host = SIMIX_host_self();
+MPI_Request smpi_mpi_isend(void* buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) {
+  MPI_Request request = smpi_isend_init(buf, count, datatype, dst, tag, comm);
 
-  hdata = xbt_new(s_smpi_host_data_t, 1);
-  SIMIX_host_set_data(host, hdata);
+  smpi_mpi_start(request);
+  return request;
+}
 
-  for (i = 0; i < smpi_global->host_count && host != smpi_global->hosts[i]; i++);
+MPI_Request smpi_irecv_init(void* buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) {
+  MPI_Request request = build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag, comm, NON_PERSISTENT | RECV);
 
-  hdata->index = i;
-  hdata->mutex = SIMIX_mutex_init();
-  hdata->cond = SIMIX_cond_init();
+  return request;
+}
 
-  hdata->pending_recv_request_queue = xbt_fifo_new();
+MPI_Request smpi_mpi_irecv(void* buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) {
+  MPI_Request request = smpi_irecv_init(buf, count, datatype, src, tag, comm);
 
-  hdata->main = SIMIX_process_self();
-  hdata->sender = SIMIX_process_create("smpi_sender",
-          smpi_sender, hdata,
-          SIMIX_host_get_name(SIMIX_host_self()), 0, NULL,
-          /*props */ NULL);
-  hdata->receiver = SIMIX_process_create("smpi_receiver",
-          smpi_receiver, hdata,
-          SIMIX_host_get_name(SIMIX_host_self()), 0, NULL,
-          /*props */ NULL);
+  smpi_mpi_start(request);
+  return request;
+}
 
-  smpi_global->main_processes[hdata->index] = SIMIX_process_self();
-  return;
+void smpi_mpi_recv(void* buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status* status) {
+  MPI_Request request;
+
+  request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
+  smpi_mpi_wait(&request, status);
 }
 
-void smpi_process_finalize()
-{
-  int i;
-  smpi_host_data_t hdata =  SIMIX_host_get_data(SIMIX_host_self());
+void smpi_mpi_send(void* buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) {
+  MPI_Request request;
 
-  i = --smpi_global->running_hosts_count;
+  request = smpi_mpi_isend(buf, count, datatype, dst, tag, comm);
+  smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
+}
 
-  SIMIX_mutex_destroy(smpi_host_mutex());
-  SIMIX_cond_destroy(smpi_host_cond());
-  xbt_fifo_free(hdata->pending_recv_request_queue);
+void smpi_mpi_sendrecv(void* sendbuf, int sendcount, MPI_Datatype sendtype, int dst, int sendtag, void* recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag, MPI_Comm comm, MPI_Status* status) {
+  MPI_Request requests[2];
+  MPI_Status stats[2];
+
+  requests[0] = smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
+  requests[1] = smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
+  smpi_mpi_startall(2, requests);
+  smpi_mpi_waitall(2, requests, stats);
+  if(status != MPI_STATUS_IGNORE) {
+    // Copy receive status
+    memcpy(status, &stats[1], sizeof(MPI_Status));
+  }
+}
 
-  if (0 >= i) {
+static void finish_wait(MPI_Request* request, MPI_Status* status) {
+  if(status != MPI_STATUS_IGNORE) {
+    status->MPI_SOURCE = (*request)->src;
+    status->MPI_TAG = (*request)->tag;
+    status->MPI_ERROR = MPI_SUCCESS;
+  }
+  print_request("finishing wait", *request);
+  if((*request)->complete == 1) {
+    SIMIX_rdv_destroy((*request)->rdv);
+  } else {
+    (*request)->match->complete = 1;
+    (*request)->match->match = MPI_REQUEST_NULL;
+  }
+  if(((*request)->flags & NON_PERSISTENT) == NON_PERSISTENT) {
+    smpi_mpi_request_free(request);
+  } else {
+    (*request)->rdv = NULL;
+    (*request)->pair = NULL;
+  }
+}
+
+int smpi_mpi_test(MPI_Request* request, MPI_Status* status) {
+  int flag = (*request)->complete;
 
-    // wake up senders/receivers
-    for (i = 0; i < smpi_global->host_count; i++) {
-      if (SIMIX_process_is_suspended(smpi_global->sender_processes[i])) {
-        SIMIX_process_resume(smpi_global->sender_processes[i]);
+  if(flag) {
+    SIMIX_communication_destroy((*request)->pair);
+    finish_wait(request, status);
+  }
+  return flag;
+}
+
+int smpi_mpi_testany(int count, MPI_Request requests[], int* index, MPI_Status* status) {
+  int i, flag;
+
+  *index = MPI_UNDEFINED;
+  flag = 0;
+  for(i = 0; i < count; i++) {
+    if(requests[i] != MPI_REQUEST_NULL && requests[i]->complete) {
+      SIMIX_communication_destroy(requests[i]->pair);
+      finish_wait(&requests[i], status);
+      *index = i;
+      flag = 1;
+      break;
+    }
+  }
+  return flag;
+}
+
+void smpi_mpi_wait(MPI_Request* request, MPI_Status* status) {
+  print_request("wait", *request);
+  // data is null if receiver waits before sender enters the rdv
+  if((*request)->complete) {
+    SIMIX_communication_destroy((*request)->pair);
+  } else {
+    SIMIX_network_wait((*request)->pair, -1.0);
+  }
+  finish_wait(request, status);
+}
+
+int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status* status) {
+  xbt_dynar_t comms;
+  int i, size, index;
+  int* map;
+
+  index = MPI_UNDEFINED;
+  if(count > 0) {
+    // First check for already completed requests
+    for(i = 0; i < count; i++) {
+      if(requests[i] != MPI_REQUEST_NULL && requests[i]->complete) {
+        index = i;
+        SIMIX_communication_destroy(requests[index]->pair); // always succeeds (but cleans the simix layer)
+        break;
+      }
+    }
+    if(index == MPI_UNDEFINED) {
+      // Otherwise, wait for a request to complete
+      comms = xbt_dynar_new(sizeof(smx_comm_t), NULL);
+      map = xbt_new(int, count);
+      size = 0;
+      DEBUG0("Wait for one of");
+      for(i = 0; i < count; i++) {
+        if(requests[i] != MPI_REQUEST_NULL && requests[i]->complete == 0) {
+          print_request("   ", requests[i]);
+          xbt_dynar_push(comms, &requests[i]->pair);
+          map[size] = i;
+          size++;
+        }
       }
-      if (SIMIX_process_is_suspended(smpi_global->receiver_processes[i])) {
-        SIMIX_process_resume(smpi_global->receiver_processes[i]);
+      if(size > 0) {
+        index = SIMIX_network_waitany(comms);
+        index = map[index];
       }
+      xbt_free(map);
+      xbt_dynar_free(&comms);
     }
+    if(index != MPI_UNDEFINED) {
+      finish_wait(&requests[index], status);
+    }
+  }
+  return index;
+}
 
-    SIMIX_mutex_destroy(smpi_mpi_global->mpi_comm_world->barrier_mutex);
-    SIMIX_cond_destroy(smpi_mpi_global->mpi_comm_world->barrier_cond);
-    xbt_free(smpi_mpi_global->mpi_comm_world);
-
-    xbt_free(smpi_mpi_global->mpi_byte);
-    xbt_free(smpi_mpi_global->mpi_int);
-    xbt_free(smpi_mpi_global->mpi_double);
+void smpi_mpi_waitall(int count, MPI_Request requests[],  MPI_Status status[]) {
+  int index;
+  MPI_Status stat;
 
-    xbt_free(smpi_mpi_global->mpi_land);
-    xbt_free(smpi_mpi_global->mpi_sum);
+  while(count > 0) {
+    index = smpi_mpi_waitany(count, requests, &stat);
+    if(index == MPI_UNDEFINED) {
+      break;
+    }
+    if(status != MPI_STATUS_IGNORE) {
+      memcpy(&status[index], &stat, sizeof(stat));
+    }
+    // FIXME: check this -v
+    // Move the last request to the found position
+    requests[index] = requests[count - 1];
+    requests[count - 1] = MPI_REQUEST_NULL;
+    count--;
+  }
+}
 
-    xbt_free(smpi_mpi_global);
+int smpi_mpi_waitsome(int incount, MPI_Request requests[], int* indices, MPI_Status status[]) {
+  int i, count;
 
+  count = 0;
+  for(i = 0; i < incount; i++) {
+    if(requests[i] != MPI_REQUEST_NULL && requests[i]->complete) {
+      SIMIX_communication_destroy(requests[i]->pair);
+      finish_wait(&requests[i], status != MPI_STATUS_IGNORE ? &status[i] : MPI_STATUS_IGNORE);
+      indices[count] = i;
+      count++;
+    }
   }
+  return count;
+}
 
+void smpi_mpi_bcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm) {
+  // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
+  nary_tree_bcast(buf, count, datatype, root, comm, 4);
 }
 
-int smpi_mpi_barrier(smpi_mpi_communicator_t comm)
-{
+void smpi_mpi_barrier(MPI_Comm comm) {
+  // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
+  nary_tree_barrier(comm, 4);
+}
 
-  SIMIX_mutex_lock(comm->barrier_mutex);
-  ++comm->barrier_count;
-  if (comm->barrier_count > comm->size) {       // only happens on second barrier...
-    comm->barrier_count = 0;
-  } else if (comm->barrier_count == comm->size) {
-    SIMIX_cond_broadcast(comm->barrier_cond);
-  }
-  while (comm->barrier_count < comm->size) {
-    SIMIX_cond_wait(comm->barrier_cond, comm->barrier_mutex);
-  }
-  SIMIX_mutex_unlock(comm->barrier_mutex);
+void smpi_mpi_gather(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) {
+  int system_tag = 666;
+  int rank, size, src, index, sendsize, recvsize;
+  MPI_Request* requests;
 
-  return MPI_SUCCESS;
+  rank = smpi_comm_rank(comm);
+  size = smpi_comm_size(comm);
+  if(rank != root) {
+    // Send buffer to root
+    smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
+  } else {
+    sendsize = smpi_datatype_size(sendtype);
+    recvsize = smpi_datatype_size(recvtype);
+    // Local copy from root
+    memcpy(&((char*)recvbuf)[root * recvcount * recvsize], sendbuf, sendcount * sendsize * sizeof(char));
+    // Receive buffers from senders
+    requests = xbt_new(MPI_Request, size - 1);
+    index = 0;
+    for(src = 0; src < size; src++) {
+      if(src != root) {
+        requests[index] = smpi_irecv_init(&((char*)recvbuf)[src * recvcount * recvsize], recvcount, recvtype, src, system_tag, comm);
+        index++;
+      }
+    }
+    // Wait for completion of irecv's.
+    smpi_mpi_startall(size - 1, requests);
+    smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
+    xbt_free(requests);
+  }
 }
 
-int smpi_mpi_isend(smpi_mpi_request_t request)
-{
-  int retval = MPI_SUCCESS;
-  int index = smpi_host_index();
+void smpi_mpi_gatherv(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int* recvcounts, int* displs, MPI_Datatype recvtype, int root, MPI_Comm comm) {
+  int system_tag = 666;
+  int rank, size, src, index, sendsize;
+  MPI_Request* requests;
 
-  if (NULL == request) {
-    retval = MPI_ERR_INTERN;
+  rank = smpi_comm_rank(comm);
+  size = smpi_comm_size(comm);
+  if(rank != root) {
+    // Send buffer to root
+    smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
   } else {
-    xbt_fifo_push(smpi_global->pending_send_request_queues[index], request);
+    sendsize = smpi_datatype_size(sendtype);
+    // Local copy from root
+    memcpy(&((char*)recvbuf)[displs[root]], sendbuf, sendcount * sendsize * sizeof(char));
+    // Receive buffers from senders
+    requests = xbt_new(MPI_Request, size - 1);
+    index = 0;
+    for(src = 0; src < size; src++) {
+      if(src != root) {
+        requests[index] = smpi_irecv_init(&((char*)recvbuf)[displs[src]], recvcounts[src], recvtype, src, system_tag, comm);
+        index++;
+      }
+    }
+    // Wait for completion of irecv's.
+    smpi_mpi_startall(size - 1, requests);
+    smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
+    xbt_free(requests);
+  }
+}
 
-    if (SIMIX_process_is_suspended(smpi_global->sender_processes[index])) {
-      SIMIX_process_resume(smpi_global->sender_processes[index]);
+void smpi_mpi_allgather(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) {
+  int system_tag = 666;
+  int rank, size, other, index, sendsize, recvsize;
+  MPI_Request* requests;
+
+  rank = smpi_comm_rank(comm);
+  size = smpi_comm_size(comm);
+  sendsize = smpi_datatype_size(sendtype);
+  recvsize = smpi_datatype_size(recvtype);
+  // Local copy from self
+  memcpy(&((char*)recvbuf)[rank * recvcount * recvsize], sendbuf, sendcount * sendsize * sizeof(char));
+  // Send/Recv buffers to/from others;
+  requests = xbt_new(MPI_Request, 2 * (size - 1));
+  index = 0;
+  for(other = 0; other < size; other++) {
+    if(other != rank) {
+      requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, comm);
+      index++;
+      requests[index] = smpi_irecv_init(&((char*)recvbuf)[other * recvcount * recvsize], recvcount, recvtype, other, system_tag, comm);
+      index++;
     }
   }
+  // Wait for completion of all comms.
+  smpi_mpi_startall(2 * (size - 1), requests);
+  smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
+  xbt_free(requests);
+}
 
-  return retval;
+void smpi_mpi_allgatherv(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int* recvcounts, int* displs, MPI_Datatype recvtype, MPI_Comm comm) {
+  int system_tag = 666;
+  int rank, size, other, index, sendsize, recvsize;
+  MPI_Request* requests;
+
+  rank = smpi_comm_rank(comm);
+  size = smpi_comm_size(comm);
+  sendsize = smpi_datatype_size(sendtype);
+  recvsize = smpi_datatype_size(recvtype);
+  // Local copy from self
+  memcpy(&((char*)recvbuf)[displs[rank]], sendbuf, sendcount * sendsize * sizeof(char));
+  // Send buffers to others;
+  requests = xbt_new(MPI_Request, 2 * (size - 1));
+  index = 0;
+  for(other = 0; other < size; other++) {
+    if(other != rank) {
+      requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, comm);
+      index++;
+      requests[index] = smpi_irecv_init(&((char*)recvbuf)[displs[other]], recvcounts[other], recvtype, other, system_tag, comm);
+      index++;
+    }
+  }
+  // Wait for completion of all comms.
+  smpi_mpi_startall(2 * (size - 1), requests);
+  smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
+  xbt_free(requests);
 }
 
-int smpi_mpi_irecv(smpi_mpi_request_t request)
-{
-  int retval = MPI_SUCCESS;
-  smpi_host_data_t hdata =  SIMIX_host_get_data(SIMIX_host_self());
+void smpi_mpi_scatter(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) {
+  int system_tag = 666;
+  int rank, size, dst, index, sendsize, recvsize;
+  MPI_Request* requests;
 
-  if (NULL == request) {
-    retval = MPI_ERR_INTERN;
+  rank = smpi_comm_rank(comm);
+  size = smpi_comm_size(comm);
+  if(rank != root) {
+    // Recv buffer from root
+    smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
   } else {
-    xbt_fifo_push(hdata->pending_recv_request_queue, request);
-
-    if (SIMIX_process_is_suspended(smpi_global->receiver_processes[hdata->index])) {
-      SIMIX_process_resume(smpi_global->receiver_processes[hdata->index]);
+    sendsize = smpi_datatype_size(sendtype);
+    recvsize = smpi_datatype_size(recvtype);
+    // Local copy from root
+    memcpy(recvbuf, &((char*)sendbuf)[root * sendcount * sendsize], recvcount * recvsize * sizeof(char));
+    // Send buffers to receivers
+    requests = xbt_new(MPI_Request, size - 1);
+    index = 0;
+    for(dst = 0; dst < size; dst++) {
+      if(dst != root) {
+        requests[index] = smpi_isend_init(&((char*)sendbuf)[dst * sendcount * sendsize], sendcount, sendtype, dst, system_tag, comm);
+        index++;
+      }
     }
+    // Wait for completion of isend's.
+    smpi_mpi_startall(size - 1, requests);
+    smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
+    xbt_free(requests);
   }
-
-  return retval;
 }
 
-int smpi_mpi_wait(smpi_mpi_request_t request, smpi_mpi_status_t * status)
-{
-  int retval = MPI_SUCCESS;
+void smpi_mpi_scatterv(void* sendbuf, int* sendcounts, int* displs, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) {
+  int system_tag = 666;
+  int rank, size, dst, index, sendsize, recvsize;
+  MPI_Request* requests;
 
-  if (NULL == request) {
-    retval = MPI_ERR_INTERN;
+  rank = smpi_comm_rank(comm);
+  size = smpi_comm_size(comm);
+  if(rank != root) {
+    // Recv buffer from root
+    smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
   } else {
-    SIMIX_mutex_lock(request->mutex);
-    while (!request->completed) {
-      SIMIX_cond_wait(request->cond, request->mutex);
+    sendsize = smpi_datatype_size(sendtype);
+    recvsize = smpi_datatype_size(recvtype);
+    // Local copy from root
+    memcpy(recvbuf, &((char*)sendbuf)[displs[root]], recvcount * recvsize * sizeof(char));
+    // Send buffers to receivers
+    requests = xbt_new(MPI_Request, size - 1);
+    index = 0;
+    for(dst = 0; dst < size; dst++) {
+      if(dst != root) {
+        requests[index] = smpi_isend_init(&((char*)sendbuf)[displs[dst]], sendcounts[dst], sendtype, dst, system_tag, comm);
+        index++;
+      }
     }
-    if (NULL != status) {
-      status->MPI_SOURCE = request->src;
-      status->MPI_TAG = request->tag;
-      status->MPI_ERROR = MPI_SUCCESS;
+    // Wait for completion of isend's.
+    smpi_mpi_startall(size - 1, requests);
+    smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
+    xbt_free(requests);
+  }
+}
+
+void smpi_mpi_reduce(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) {
+  int system_tag = 666;
+  int rank, size, src, index, datasize;
+  MPI_Request* requests;
+  void** tmpbufs;
+
+  rank = smpi_comm_rank(comm);
+  size = smpi_comm_size(comm);
+  if(rank != root) {
+    // Send buffer to root
+    smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm);
+  } else {
+    datasize = smpi_datatype_size(datatype);
+    // Local copy from root
+    memcpy(recvbuf, sendbuf, count * datasize * sizeof(char));
+    // Receive buffers from senders
+    //TODO: make a MPI_barrier here ?
+    requests = xbt_new(MPI_Request, size - 1);
+    tmpbufs = xbt_new(void*, size - 1);
+    index = 0;
+    for(src = 0; src < size; src++) {
+      if(src != root) {
+        tmpbufs[index] = xbt_malloc(count * datasize);
+        requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, src, system_tag, comm);
+        index++;
+      }
     }
-    SIMIX_mutex_unlock(request->mutex);
+    // Wait for completion of irecv's.
+    smpi_mpi_startall(size - 1, requests);
+    for(src = 0; src < size - 1; src++) {
+      index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
+      if(index == MPI_UNDEFINED) {
+        break;
+      }
+      smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+    }
+    for(index = 0; index < size - 1; index++) {
+      xbt_free(tmpbufs[index]);
+    }
+    xbt_free(tmpbufs);
+    xbt_free(requests);
   }
+}
 
-  return retval;
+void smpi_mpi_allreduce(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) {
+  smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
+  smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
+
+/*
+FIXME: buggy implementation
+
+  int system_tag = 666;
+  int rank, size, other, index, datasize;
+  MPI_Request* requests;
+  void** tmpbufs;
+
+  rank = smpi_comm_rank(comm);
+  size = smpi_comm_size(comm);
+  datasize = smpi_datatype_size(datatype);
+  // Local copy from self
+  memcpy(recvbuf, sendbuf, count * datasize * sizeof(char));
+  // Send/Recv buffers to/from others;
+  //TODO: make a MPI_barrier here ?
+  requests = xbt_new(MPI_Request, 2 * (size - 1));
+  tmpbufs = xbt_new(void*, size - 1);
+  index = 0;
+  for(other = 0; other < size; other++) {
+    if(other != rank) {
+      tmpbufs[index / 2] = xbt_malloc(count * datasize);
+      requests[index] = smpi_mpi_isend(sendbuf, count, datatype, other, system_tag, comm);
+      requests[index + 1] = smpi_mpi_irecv(tmpbufs[index / 2], count, datatype, other, system_tag, comm);
+      index += 2;
+    }
+  }
+  // Wait for completion of all comms.
+  for(other = 0; other < 2 * (size - 1); other++) {
+    index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
+    if(index == MPI_UNDEFINED) {
+      break;
+    }
+    if((index & 1) == 1) {
+      // Request is odd: it's a irecv
+      smpi_op_apply(op, tmpbufs[index / 2], recvbuf, &count, &datatype);
+    }
+  }
+  for(index = 0; index < size - 1; index++) {
+    xbt_free(tmpbufs[index]);
+  }
+  xbt_free(tmpbufs);
+  xbt_free(requests);
+*/
+}
+
+void smpi_mpi_scan(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) {
+  int system_tag = 666;
+  int rank, size, other, index, datasize;
+  int total;
+  MPI_Request* requests;
+  void** tmpbufs;
+
+  rank = smpi_comm_rank(comm);
+  size = smpi_comm_size(comm);
+  datasize = smpi_datatype_size(datatype);
+  // Local copy from self
+  memcpy(recvbuf, sendbuf, count * datasize * sizeof(char));
+  // Send/Recv buffers to/from others;
+  total = rank + (size - (rank + 1));
+  requests = xbt_new(MPI_Request, total);
+  tmpbufs = xbt_new(void*, rank);
+  index = 0;
+  for(other = 0; other < rank; other++) {
+    tmpbufs[index] = xbt_malloc(count * datasize);
+    requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
+    index++;
+  }
+  for(other = rank + 1; other < size; other++) {
+    requests[index] = smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
+    index++;
+  }
+  // Wait for completion of all comms.
+  smpi_mpi_startall(size - 1, requests);
+  for(other = 0; other < total; other++) {
+    index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
+    if(index == MPI_UNDEFINED) {
+      break;
+    }
+    if(index < rank) {
+      // #Request is below rank: it's a irecv
+      smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+    }
+  }
+  for(index = 0; index < size - 1; index++) {
+    xbt_free(tmpbufs[index]);
+  }
+  xbt_free(tmpbufs);
+  xbt_free(requests);
 }