X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/cc375ad134c19925e1e343bae932be21cf8a8caf..db17323a11a3089f492c2cae1be7913673689d5c:/src/smpi/smpi_base.c

diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c
index 2c99064110..88e93074d0 100644
--- a/src/smpi/smpi_base.c
+++ b/src/smpi/smpi_base.c
@@ -12,7 +12,7 @@
 #include "simix/smx_private.h"
 #include "surf/surf.h"
 #include "simgrid/sg_config.h"
-
+#include "colls/colls.h"
 
 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
 
@@ -25,7 +25,7 @@ static int match_recv(void* a, void* b, smx_action_t ignored) {
   xbt_assert(ref, "Cannot match recv against null reference");
   xbt_assert(req, "Cannot match recv against null request");
   if((ref->src == MPI_ANY_SOURCE || req->src == ref->src)
-    && (ref->tag == MPI_ANY_TAG || req->tag == ref->tag)){
+    && ((ref->tag == MPI_ANY_TAG && req->tag >=0) || req->tag == ref->tag)){
     //we match, we can transfer some values
     // FIXME : move this to the copy function ?
     if(ref->src == MPI_ANY_SOURCE)ref->real_src = req->src;
@@ -34,6 +34,7 @@ static int match_recv(void* a, void* b, smx_action_t ignored) {
     if(req->detached==1){
         ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
     }
+    XBT_DEBUG("match succeeded");
     return 1;
   }else return 0;
 }
@@ -46,7 +47,7 @@ static int match_send(void* a, void* b,smx_action_t ignored) {
    xbt_assert(req, "Cannot match send against null request");
 
    if((req->src == MPI_ANY_SOURCE || req->src == ref->src)
-             && (req->tag == MPI_ANY_TAG || req->tag == ref->tag))
+             && ((req->tag == MPI_ANY_TAG && ref->tag >=0)|| req->tag == ref->tag))
    {
      if(req->src == MPI_ANY_SOURCE)req->real_src = ref->src;
      if(req->tag == MPI_ANY_TAG)req->real_tag = ref->tag;
@@ -54,11 +55,134 @@ static int match_send(void* a, void* b,smx_action_t ignored) {
      if(ref->detached==1){
          req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
      }
-
+    XBT_DEBUG("match succeeded");
      return 1;
    } else return 0;
 }
 
+
+typedef struct s_smpi_factor *smpi_factor_t;
+typedef struct s_smpi_factor {
+  long factor;
+  int nb_values;
+  double values[4];//arbitrary set to 4
+} s_smpi_factor_t;
+xbt_dynar_t smpi_os_values = NULL;
+xbt_dynar_t smpi_or_values = NULL;
+xbt_dynar_t smpi_ois_values = NULL;
+
+// Methods used to parse and store the values for timing injections in smpi
+// These are taken from surf/network.c and generalized to have more factors
+// These methods should be merged with those in surf/network.c (moved somewhere in xbt ?)
+
+static int factor_cmp(const void *pa, const void *pb)
+{
+  return (((s_smpi_factor_t*)pa)->factor > ((s_smpi_factor_t*)pb)->factor);
+}
+
+
+static xbt_dynar_t parse_factor(const char *smpi_coef_string)
+{
+  char *value = NULL;
+  unsigned int iter = 0;
+  s_smpi_factor_t fact;
+  int i=0;
+  xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL;
+
+  smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL);
+  radical_elements = xbt_str_split(smpi_coef_string, ";");
+  xbt_dynar_foreach(radical_elements, iter, value) {
+    fact.nb_values=0;
+    radical_elements2 = xbt_str_split(value, ":");
+    if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
+      xbt_die("Malformed radical for smpi factor!");
+    for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
+        if (i==0){
+           fact.factor = atol(xbt_dynar_get_as(radical_elements2, i, char *));
+        }else{
+           fact.values[fact.nb_values] = atof(xbt_dynar_get_as(radical_elements2, i, char *));
+           fact.nb_values++;
+        }
+    }
+
+    xbt_dynar_push_as(smpi_factor, s_smpi_factor_t, fact);
+    XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
+    xbt_dynar_free(&radical_elements2);
+  }
+  xbt_dynar_free(&radical_elements);
+  iter=0;
+  xbt_dynar_sort(smpi_factor, &factor_cmp);
+  xbt_dynar_foreach(smpi_factor, iter, fact) {
+    XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
+  }
+  return smpi_factor;
+}
+
+static double smpi_os(double size)
+{
+  if (!smpi_os_values) {
+    smpi_os_values = parse_factor(sg_cfg_get_string("smpi/os"));
+    smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
+  }
+  unsigned int iter = 0;
+  s_smpi_factor_t fact;
+  double current=0.0;
+  xbt_dynar_foreach(smpi_os_values, iter, fact) {
+    if (size <= fact.factor) {
+        XBT_DEBUG("os : %lf <= %ld return %f", size, fact.factor, current);
+      return current;
+    }else{
+      current=fact.values[0]+fact.values[1]*size;
+    }
+  }
+  XBT_DEBUG("os : %lf > %ld return %f", size, fact.factor, current);
+
+  return current;
+}
+
+static double smpi_ois(double size)
+{
+  if (!smpi_ois_values) {
+    smpi_ois_values = parse_factor(sg_cfg_get_string("smpi/ois"));
+    smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
+  }
+  unsigned int iter = 0;
+  s_smpi_factor_t fact;
+  double current=0.0;
+  xbt_dynar_foreach(smpi_ois_values, iter, fact) {
+    if (size <= fact.factor) {
+        XBT_DEBUG("ois : %lf <= %ld return %f", size, fact.factor, current);
+      return current;
+    }else{
+      current=fact.values[0]+fact.values[1]*size;
+    }
+  }
+  XBT_DEBUG("ois : %lf > %ld return %f", size, fact.factor, current);
+
+  return current;
+}
+
+static double smpi_or(double size)
+{
+  if (!smpi_or_values) {
+    smpi_or_values = parse_factor(sg_cfg_get_string("smpi/or"));
+    smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
+  }
+  unsigned int iter = 0;
+  s_smpi_factor_t fact;
+  double current=0.0;
+  xbt_dynar_foreach(smpi_or_values, iter, fact) {
+    if (size <= fact.factor) {
+        XBT_DEBUG("or : %lf <= %ld return %f", size, fact.factor, current);
+      return current;
+    }else
+      current=fact.values[0]+fact.values[1]*size;
+  }
+  XBT_DEBUG("or : %lf > %ld return %f", size, fact.factor, current);
+
+  return current;
+}
+
 static MPI_Request build_request(void *buf, int count,
                                  MPI_Datatype datatype, int src, int dst,
                                  int tag, MPI_Comm comm, unsigned flags)
@@ -74,7 +198,7 @@ static MPI_Request build_request(void *buf, int count,
   if(datatype->has_subtype == 1){
     // This part handles the problem of non-contiguous memory
     old_buf = buf;
-    buf = xbt_malloc(count*smpi_datatype_size(datatype));
+    buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
     if (flags & SEND) {
       subtype->serialize(old_buf, buf, count, datatype->substruct);
     }
@@ -111,11 +235,13 @@ static MPI_Request build_request(void *buf, int count,
 }
 
 
-void smpi_empty_status(MPI_Status * status) {
+void smpi_empty_status(MPI_Status * status)
+{
   if(status != MPI_STATUS_IGNORE) {
-      status->MPI_SOURCE=MPI_ANY_SOURCE;
-      status->MPI_TAG=MPI_ANY_TAG;
-      status->count=0;
+    status->MPI_SOURCE = MPI_ANY_SOURCE;
+    status->MPI_TAG = MPI_ANY_TAG;
+    status->MPI_ERROR = MPI_SUCCESS;
+    status->count=0;
   }
 }
 
@@ -159,8 +285,18 @@ 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);
+    build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
+                  comm, PERSISTENT | SEND | PREPARED);
+  request->refcount++;
+  return request;
+}
+
+MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
+                               int dst, int tag, MPI_Comm comm)
+{
+  MPI_Request request =
+    build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
+                  comm, PERSISTENT | SSEND | SEND | PREPARED);
   request->refcount++;
   return request;
 }
@@ -169,8 +305,8 @@ 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);
+    build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
+                  comm, PERSISTENT | RECV | PREPARED);
   request->refcount++;
   return request;
 }
@@ -181,6 +317,7 @@ void smpi_mpi_start(MPI_Request request)
 
   xbt_assert(!request->action,
              "Cannot (re)start a non-finished communication");
+  if(request->flags & PREPARED)request->flags &= ~PREPARED;
   if(request->flags & RECV) {
     print_request("New recv", request);
     if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
@@ -190,10 +327,27 @@ void smpi_mpi_start(MPI_Request request)
     // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
     request->real_size=request->size;
     smpi_datatype_use(request->old_type);
+    smpi_comm_use(request->comm);
     request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request);
+
+    //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
+    double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
+    if(sleeptime!=0.0){
+        simcall_process_sleep(sleeptime);
+        XBT_DEBUG("receiving size of %zu : sleep %lf ", request->size, smpi_or(request->size));
+    }
+
   } else {
 
-    int receiver = smpi_group_index(smpi_comm_group(request->comm), request->dst);
+
+    int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
+
+    #ifdef HAVE_TRACING
+      int rank = smpi_process_index();
+      if (TRACE_smpi_view_internals()) {
+        TRACE_smpi_send(rank, rank, receiver);
+      }
+    #endif
 /*    if(receiver == MPI_UNDEFINED) {*/
 /*      XBT_WARN("Trying to send a message to a wrong rank");*/
 /*      return;*/
@@ -205,22 +359,36 @@ void smpi_mpi_start(MPI_Request request)
       XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
       mailbox = smpi_process_remote_mailbox(receiver);
     }
-    if (request->size < 64*1024 ) { //(FIXME: this limit should be configurable)
+    if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
       void *oldbuf = NULL;
       request->detached = 1;
       request->refcount++;
       if(request->old_type->has_subtype == 0){
         oldbuf = request->buf;
-        if (oldbuf){
+        if (oldbuf && request->size!=0){
           request->buf = xbt_malloc(request->size);
           memcpy(request->buf,oldbuf,request->size);
         }
       }
       XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
     }
+
     // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
     request->real_size=request->size;
     smpi_datatype_use(request->old_type);
+    smpi_comm_use(request->comm);
+
+    //if we are giving back the control to the user without waiting for completion, we have to inject timings
+    double sleeptime =0.0;
+    if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
+      //isend and send timings may be different
+      sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
+    }
+
+    if(sleeptime!=0.0){
+        simcall_process_sleep(sleeptime);
+        XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size));
+    }
 
     request->action =
       simcall_comm_isend(mailbox, request->size, -1.0,
@@ -235,6 +403,7 @@ void smpi_mpi_start(MPI_Request request)
     /* FIXME: detached sends are not traceable (request->action == NULL) */
     if (request->action)
       simcall_set_category(request->action, TRACE_internal_smpi_get_category());
+
 #endif
 
   }
@@ -252,14 +421,17 @@ void smpi_mpi_startall(int count, MPI_Request * requests)
 
 void smpi_mpi_request_free(MPI_Request * request)
 {
-
   if((*request) != MPI_REQUEST_NULL){
     (*request)->refcount--;
     if((*request)->refcount<0) xbt_die("wrong refcount");
 
     if((*request)->refcount==0){
+        print_request("Destroying", (*request));
         xbt_free(*request);
         *request = MPI_REQUEST_NULL;
+    }else{
+        print_request("Decrementing", (*request));
+
     }
   }else{
       xbt_die("freeing an already free request");
@@ -270,8 +442,8 @@ 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);
+    build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
+                  comm, NON_PERSISTENT | SEND | PREPARED);
 
   return request;
 }
@@ -280,19 +452,31 @@ MPI_Request smpi_mpi_isend(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);
+      build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
+                    comm, NON_PERSISTENT | ISEND | SEND);
+
+  smpi_mpi_start(request);
+  return request;
+}
 
+MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
+                           int dst, int tag, MPI_Comm comm)
+{
+  MPI_Request request =
+      build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
+                    comm, NON_PERSISTENT | ISEND | SSEND | SEND);
   smpi_mpi_start(request);
   return request;
 }
 
+
+
 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);
+    build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
+                  comm, NON_PERSISTENT | RECV | PREPARED);
   return request;
 }
 
@@ -300,7 +484,7 @@ MPI_Request smpi_mpi_irecv(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,
+      build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
                     comm, NON_PERSISTENT | RECV);
 
   smpi_mpi_start(request);
@@ -320,12 +504,25 @@ void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
                    int tag, MPI_Comm comm)
 {
-  MPI_Request request;
-  request = smpi_mpi_isend(buf, count, datatype, dst, tag, comm);
+  MPI_Request request =
+      build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
+                    comm, NON_PERSISTENT | SEND);
+  smpi_mpi_start(request);
   smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
 
 }
 
+void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
+                           int dst, int tag, MPI_Comm comm)
+{
+  MPI_Request request =
+      build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
+                    comm, NON_PERSISTENT | SSEND | SEND);
+
+  smpi_mpi_start(request);
+  smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
+}
+
 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,
@@ -333,7 +530,12 @@ void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
 {
   MPI_Request requests[2];
   MPI_Status stats[2];
-
+  int myid=smpi_process_index();
+  if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
+      smpi_datatype_copy(sendbuf, sendcount, sendtype,
+                                     recvbuf, recvcount, recvtype);
+      return;
+  }
   requests[0] =
     smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
   requests[1] =
@@ -342,7 +544,7 @@ void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   smpi_mpi_waitall(2, requests, stats);
   if(status != MPI_STATUS_IGNORE) {
     // Copy receive status
-    memcpy(status, &stats[1], sizeof(MPI_Status));
+    *status = stats[1];
   }
 }
 
@@ -354,13 +556,15 @@ int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
 static void finish_wait(MPI_Request * request, MPI_Status * status)
 {
   MPI_Request req = *request;
-  if(!(req->detached && req->flags & SEND)){
-    if(status != MPI_STATUS_IGNORE) {
-      status->MPI_SOURCE = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
+  if(status != MPI_STATUS_IGNORE)
+    smpi_empty_status(status);
+
+  if(!(req->detached && req->flags & SEND) && !(req->flags & PREPARED)){
+     if(status != MPI_STATUS_IGNORE) {
+      int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
+      status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
       status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
-      if(req->truncated)
-      status->MPI_ERROR = MPI_ERR_TRUNCATE;
-      else status->MPI_ERROR = MPI_SUCCESS ;
+      status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
       // this handles the case were size in receive differs from size in send
       // FIXME: really this should just contain the count of receive-type blocks,
       // right?
@@ -379,9 +583,21 @@ static void finish_wait(MPI_Request * request, MPI_Status * status)
       }
       if(req->detached == 0) free(req->buf);
     }
+    smpi_comm_unuse(req->comm);
     smpi_datatype_unuse(datatype);
+
   }
 
+#ifdef HAVE_TRACING
+    if (TRACE_smpi_view_internals()) {
+      if(req->flags & RECV){
+        int rank = smpi_process_index();
+        int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
+        TRACE_smpi_recv(rank, src_traced, rank);
+      }
+    }
+#endif
+
   if(req->detached_sender!=NULL){
     smpi_mpi_request_free(&(req->detached_sender));
   }
@@ -402,8 +618,8 @@ int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
   else
     flag = simcall_comm_test((*request)->action);
   if(flag) {
-    (*request)->refcount++;
     finish_wait(request, status);
+    request=MPI_REQUEST_NULL;
   }else{
     smpi_empty_status(status);
   }
@@ -419,33 +635,31 @@ int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
 
   *index = MPI_UNDEFINED;
   flag = 0;
-  if(count > 0) {
-    comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
-    map = xbt_new(int, count);
-    size = 0;
-    for(i = 0; i < count; i++) {
-      if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) {
-         xbt_dynar_push(comms, &requests[i]->action);
-         map[size] = i;
-         size++;
-      }
+  comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
+  map = xbt_new(int, count);
+  size = 0;
+  for(i = 0; i < count; i++) {
+    if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) {
+       xbt_dynar_push(comms, &requests[i]->action);
+       map[size] = i;
+       size++;
     }
-    if(size > 0) {
-      i = simcall_comm_testany(comms);
-      // not MPI_UNDEFINED, as this is a simix return code
-      if(i != -1) {
-        *index = map[i];
-        finish_wait(&requests[*index], status);
-        flag = 1;
-      }
-    }else{
-        //all requests are null or inactive, return true
-        flag=1;
-        smpi_empty_status(status);
+  }
+  if(size > 0) {
+    i = simcall_comm_testany(comms);
+    // not MPI_UNDEFINED, as this is a simix return code
+    if(i != -1) {
+      *index = map[i];
+      finish_wait(&requests[*index], status);
+      flag = 1;
     }
-    xbt_free(map);
-    xbt_dynar_free(&comms);
+  }else{
+      //all requests are null or inactive, return true
+      flag=1;
+      smpi_empty_status(status);
   }
+  xbt_free(map);
+  xbt_dynar_free(&comms);
 
   return flag;
 }
@@ -467,7 +681,7 @@ int smpi_mpi_testall(int count, MPI_Request requests[],
       smpi_empty_status(pstat);
     }
     if(status != MPI_STATUSES_IGNORE) {
-      memcpy(&status[i], pstat, sizeof(*pstat));
+      status[i] = *pstat;
     }
   }
   return flag;
@@ -480,16 +694,21 @@ void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
   while(flag==0){
     smpi_mpi_iprobe(source, tag, comm, &flag, status);
     XBT_DEBUG("Busy Waiting on probing : %d", flag);
-    if(!flag) {
-      simcall_process_sleep(0.0001);
-    }
   }
 }
 
 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
-  MPI_Request request =build_request(NULL, 0, MPI_CHAR, source, smpi_comm_rank(comm), tag,
+
+  MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag,
             comm, NON_PERSISTENT | RECV);
 
+  //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
+  double sleeptime= sg_cfg_get_double("smpi/iprobe");
+  //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
+  static int nsleeps = 1;
+
+  simcall_process_sleep(sleeptime);
+
   // behave like a receive, but don't do it
   smx_rdv_t mailbox;
 
@@ -509,14 +728,18 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   if(request->action){
     MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
     *flag = 1;
-    if(status != MPI_STATUS_IGNORE) {
-      status->MPI_SOURCE = req->src;
+    if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
+      status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
       status->MPI_TAG = req->tag;
       status->MPI_ERROR = MPI_SUCCESS;
       status->count = req->real_size;
     }
+    nsleeps=1;//reset the number of sleeps we will do next time
+  }
+  else {
+      *flag = 0;
+      nsleeps++;
   }
-  else *flag = 0;
   smpi_mpi_request_free(&request);
 
   return;
@@ -589,33 +812,32 @@ int smpi_mpi_waitall(int count, MPI_Request requests[],
   int  index, c;
   MPI_Status stat;
   MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
-  int retvalue=MPI_SUCCESS;
+  int retvalue = MPI_SUCCESS;
   //tag invalid requests in the set
-  for(c = 0; c < count; c++) {
-    if(requests[c]==MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ){
-      if(status != MPI_STATUSES_IGNORE)
+  if (status != MPI_STATUSES_IGNORE) {
+    for (c = 0; c < count; c++) {
+      if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL) {
         smpi_empty_status(&status[c]);
-    }else if(requests[c]->src == MPI_PROC_NULL ){
-      if(status != MPI_STATUSES_IGNORE) {
+      } else if (requests[c]->src == MPI_PROC_NULL) {
         smpi_empty_status(&status[c]);
-        status[c].MPI_SOURCE=MPI_PROC_NULL;
+        status[c].MPI_SOURCE = MPI_PROC_NULL;
       }
     }
   }
   for(c = 0; c < count; c++) {
-      if(MC_is_active()) {
-        smpi_mpi_wait(&requests[c], pstat);
-        index = c;
-      } else {
-        index = smpi_mpi_waitany(count, requests, pstat);
-        if(index == MPI_UNDEFINED) {
-          break;
-       }
-      if(status != MPI_STATUSES_IGNORE) {
-        memcpy(&status[index], pstat, sizeof(*pstat));
-        if(status[index].MPI_ERROR==MPI_ERR_TRUNCATE)retvalue=MPI_ERR_IN_STATUS;
-
-      }
+    if (MC_is_active()) {
+      smpi_mpi_wait(&requests[c], pstat);
+      index = c;
+    } else {
+      index = smpi_mpi_waitany(count, requests, pstat);
+      if (index == MPI_UNDEFINED)
+        break;
+      requests[index]=MPI_REQUEST_NULL;
+    }
+    if (status != MPI_STATUSES_IGNORE) {
+      status[index] = *pstat;
+      if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
+        retvalue = MPI_ERR_IN_STATUS;
     }
   }
 
@@ -637,8 +859,9 @@ int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
       indices[count] = index;
       count++;
       if(status != MPI_STATUSES_IGNORE) {
-        memcpy(&status[index], pstat, sizeof(*pstat));
+        status[index] = *pstat;
       }
+     requests[index]=MPI_REQUEST_NULL;
     }else{
       return MPI_UNDEFINED;
     }
@@ -661,8 +884,10 @@ int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
          indices[count] = i;
          count++;
          if(status != MPI_STATUSES_IGNORE) {
-            memcpy(&status[i], pstat, sizeof(*pstat));
+           status[i] = *pstat;
          }
+         requests[i]=MPI_REQUEST_NULL;
+
       }
     }else{
       count_dead++;
@@ -689,7 +914,7 @@ 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 system_tag = COLL_TAG_GATHER;
   int rank, size, src, index;
   MPI_Aint lb = 0, recvext = 0;
   MPI_Request *requests;
@@ -723,11 +948,36 @@ void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
   }
 }
 
+
+void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
+                       MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
+{
+    int i, size, count;
+    int *displs;
+    int rank = smpi_process_index();
+    void *tmpbuf;
+
+    /* arbitrarily choose root as rank 0 */
+    size = smpi_comm_size(comm);
+    count = 0;
+    displs = xbt_new(int, size);
+    for (i = 0; i < size; i++) {
+      displs[i] = count;
+      count += recvcounts[i];
+    }
+    tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
+    mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
+    smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
+                      recvcounts[rank], datatype, 0, comm);
+    xbt_free(displs);
+    xbt_free(tmpbuf);
+}
+
 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 system_tag = COLL_TAG_GATHERV;
   int rank, size, src, index;
   MPI_Aint lb = 0, recvext = 0;
   MPI_Request *requests;
@@ -767,7 +1017,7 @@ void smpi_mpi_allgather(void *sendbuf, int sendcount,
                         int recvcount, MPI_Datatype recvtype,
                         MPI_Comm comm)
 {
-  int system_tag = 666;
+  int system_tag = COLL_TAG_ALLGATHER;
   int rank, size, other, index;
   MPI_Aint lb = 0, recvext = 0;
   MPI_Request *requests;
@@ -806,7 +1056,7 @@ void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
                          int *recvcounts, int *displs,
                          MPI_Datatype recvtype, MPI_Comm comm)
 {
-  int system_tag = 666;
+  int system_tag = COLL_TAG_ALLGATHERV;
   int rank, size, other, index;
   MPI_Aint lb = 0, recvext = 0;
   MPI_Request *requests;
@@ -844,7 +1094,7 @@ 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 system_tag = COLL_TAG_SCATTER;
   int rank, size, dst, index;
   MPI_Aint lb = 0, sendext = 0;
   MPI_Request *requests;
@@ -859,8 +1109,10 @@ void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
     // FIXME: check for errors
     smpi_datatype_extent(sendtype, &lb, &sendext);
     // Local copy from root
-    smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
-                       sendcount, sendtype, recvbuf, recvcount, recvtype);
+    if(recvbuf!=MPI_IN_PLACE){
+        smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
+                           sendcount, sendtype, recvbuf, recvcount, recvtype);
+    }
     // Send buffers to receivers
     requests = xbt_new(MPI_Request, size - 1);
     index = 0;
@@ -883,7 +1135,7 @@ 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 system_tag = COLL_TAG_SCATTERV;
   int rank, size, dst, index;
   MPI_Aint lb = 0, sendext = 0;
   MPI_Request *requests;
@@ -898,8 +1150,10 @@ void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
     // FIXME: check for errors
     smpi_datatype_extent(sendtype, &lb, &sendext);
     // Local copy from root
-    smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
+    if(recvbuf!=MPI_IN_PLACE){
+      smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
                        sendtype, recvbuf, recvcount, recvtype);
+    }
     // Send buffers to receivers
     requests = xbt_new(MPI_Request, size - 1);
     index = 0;
@@ -922,23 +1176,37 @@ 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 system_tag = COLL_TAG_REDUCE;
   int rank, size, src, index;
   MPI_Aint lb = 0, dataext = 0;
   MPI_Request *requests;
   void **tmpbufs;
 
+
+  char* sendtmpbuf = (char*) sendbuf;
+  if( sendbuf == MPI_IN_PLACE ) {
+    sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
+    smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
+  }
+
   rank = smpi_comm_rank(comm);
   size = smpi_comm_size(comm);
+  //non commutative case, use a working algo from openmpi
+  if(!smpi_op_is_commute(op)){
+    smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
+                     datatype, op, root, comm);
+    return;
+  }
+  
   if(rank != root) {
     // Send buffer to root
-    smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm);
+    smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
   } else {
     // FIXME: check for errors
     smpi_datatype_extent(datatype, &lb, &dataext);
     // Local copy from root
-    if (sendbuf && recvbuf)
-      smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
+    if (sendtmpbuf && recvbuf)
+      smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
     // Receive buffers from senders
     //TODO: make a MPI_barrier here ?
     requests = xbt_new(MPI_Request, size - 1);
@@ -971,6 +1239,10 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
     }
     xbt_free(tmpbufs);
     xbt_free(requests);
+
+    if( sendbuf == MPI_IN_PLACE ) {
+      xbt_free(sendtmpbuf);
+    }
   }
 }
 
@@ -984,7 +1256,7 @@ void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
                    MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
 {
-  int system_tag = 666;
+  int system_tag = -888;
   int rank, size, other, index;
   MPI_Aint lb = 0, dataext = 0;
   MPI_Request *requests;
@@ -1019,14 +1291,94 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
   }
   // Wait for completion of all comms.
   smpi_mpi_startall(size - 1, requests);
-  for(other = 0; other < size - 1; other++) {
-    index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
-    if(index == MPI_UNDEFINED) {
-      break;
+
+  if(smpi_op_is_commute(op)){
+    for(other = 0; other < size - 1; 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);
+      }
+    }
+  }else{
+    //non commutative case, wait in order
+    for(other = 0; other < size - 1; other++) {
+      smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
+      if(index < rank) {
+        smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
+      }
     }
-    if(index < rank) {
-      // #Request is below rank: it's a irecv
-      smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+  }
+  for(index = 0; index < rank; index++) {
+    xbt_free(tmpbufs[index]);
+  }
+  xbt_free(tmpbufs);
+  xbt_free(requests);
+}
+
+void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
+                   MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
+{
+  int system_tag = -888;
+  int rank, size, other, index;
+  MPI_Aint lb = 0, dataext = 0;
+  MPI_Request *requests;
+  void **tmpbufs;
+  int recvbuf_is_empty=1;
+  rank = smpi_comm_rank(comm);
+  size = smpi_comm_size(comm);
+
+  // FIXME: check for errors
+  smpi_datatype_extent(datatype, &lb, &dataext);
+
+  // Send/Recv buffers to/from others;
+  requests = xbt_new(MPI_Request, size - 1);
+  tmpbufs = xbt_new(void *, rank);
+  index = 0;
+  for(other = 0; other < rank; other++) {
+    // FIXME: possibly overkill we we have contiguous/noncontiguous data
+    // mapping...
+    tmpbufs[index] = xbt_malloc(count * dataext);
+    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);
+  if(smpi_op_is_commute(op)){
+    for(other = 0; other < size - 1; other++) {
+      index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
+      if(index == MPI_UNDEFINED) {
+        break;
+      }
+      if(index < rank) {
+        if(recvbuf_is_empty){
+          smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
+          recvbuf_is_empty=0;
+        }else
+        // #Request is below rank: it's a irecv
+        smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+      }
+    }
+  }else{
+    //non commutative case, wait in order
+    for(other = 0; other < size - 1; other++) {
+      smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
+      if(index < rank) {
+          if(recvbuf_is_empty){
+            smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
+            recvbuf_is_empty=0;
+          }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
+      }
     }
   }
   for(index = 0; index < rank; index++) {