X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/b8d73e39d7ad879522cdb542e4ed69e2077a6f9e..8a4ded6ee6330d304e1d3f69100eb7ec55c5fd9c:/src/smpi/smpi_base.c

diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c
index 5362b82ff6..8398a13f0d 100644
--- a/src/smpi/smpi_base.c
+++ b/src/smpi/smpi_base.c
@@ -71,6 +71,13 @@ xbt_dynar_t smpi_os_values = NULL;
 xbt_dynar_t smpi_or_values = NULL;
 xbt_dynar_t smpi_ois_values = NULL;
 
+double smpi_wtime_sleep = 0.0;
+double smpi_iprobe_sleep = 1e-4;
+double smpi_test_sleep = 1e-4;
+
+xbt_dict_t smpi_keyvals = NULL;
+int keyval_id=MPI_TAG_UB+1;//avoid collisions
+
 // 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 ?)
@@ -185,6 +192,20 @@ static double smpi_or(double size)
   return current;
 }
 
+double smpi_mpi_wtime(){
+  double time;
+  if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_process_get_sampling()) {
+    smpi_bench_end();
+    time = SIMIX_get_clock();
+    //to avoid deadlocks if called too many times
+    if(smpi_wtime_sleep > 0) simcall_process_sleep(smpi_wtime_sleep);
+    smpi_bench_begin();
+  } else {
+    time = SIMIX_get_clock();
+  }
+  return time;
+}
+
 static MPI_Request build_request(void *buf, int count,
                                  MPI_Datatype datatype, int src, int dst,
                                  int tag, MPI_Comm comm, unsigned flags)
@@ -197,11 +218,11 @@ static MPI_Request build_request(void *buf, int count,
 
   s_smpi_subtype_t *subtype = datatype->substruct;
 
-  if(datatype->has_subtype == 1){
+  if(((flags & RECV) && (flags & ACCUMULATE)) || (datatype->has_subtype == 1)){
     // This part handles the problem of non-contiguous memory
     old_buf = buf;
     buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
-    if (flags & SEND) {
+    if ((datatype->has_subtype == 1) && (flags & SEND)) {
       subtype->serialize(old_buf, buf, count, datatype->substruct);
     }
   }
@@ -230,7 +251,7 @@ static MPI_Request build_request(void *buf, int count,
     request->refcount = 1;
   else
     request->refcount = 0;
-
+  request->op = MPI_REPLACE;
 #ifdef HAVE_TRACING
   request->send = 0;
   request->recv = 0;
@@ -260,8 +281,8 @@ void smpi_action_trace_run(char *path)
   action_fp=NULL;
   if (path) {
     action_fp = fopen(path, "r");
-    xbt_assert(action_fp != NULL, "Cannot open %s: %s", path,
-               strerror(errno));
+    if (action_fp == NULL)
+      xbt_die("Cannot open %s: %s", path, strerror(errno));
   }
 
   if (!xbt_dict_is_empty(action_queues)) {
@@ -317,7 +338,7 @@ MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
 void smpi_mpi_start(MPI_Request request)
 {
   smx_rdv_t mailbox;
-
+  
   xbt_assert(!request->action, "Cannot (re)start a non-finished communication");
   request->flags &= ~PREPARED;
   request->flags &= ~FINISHED;
@@ -325,89 +346,145 @@ void smpi_mpi_start(MPI_Request request)
 
   if (request->flags & RECV) {
     print_request("New recv", request);
-    //FIXME: if receive is posted with a large size, but send is smaller, mailboxes may not match !
-    if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
+    
+    if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")){
+    //We have to check both mailboxes (because SSEND messages are sent to the large mbox). begin with the more appropriate one : the small one.
       mailbox = smpi_process_mailbox_small();
-    else
-      mailbox = smpi_process_mailbox();
+      XBT_DEBUG("Is there a corresponding send already posted the small mailbox %p (in case of SSEND)?", mailbox);
+      smx_action_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
+    
+      if(action ==NULL){
+        mailbox = smpi_process_mailbox();
+        XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
+        action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
+        if(action ==NULL){
+          XBT_DEBUG("Still notching, switch back to the small mailbox : %p", mailbox);
+          mailbox = smpi_process_mailbox_small();
+          }
+      }else{
+        XBT_DEBUG("yes there was something for us in the large mailbox");
+      }
+    }else{
+      mailbox = smpi_process_mailbox_small();
+      XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
+    smx_action_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
+    
+      if(action ==NULL){
+        XBT_DEBUG("No, nothing in the permanent receive mailbox");
+        mailbox = smpi_process_mailbox();
+      }else{
+        XBT_DEBUG("yes there was something for us in the small mailbox");
+      }
+    }
+
+    //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 %f ", request->size, smpi_or(request->size));
+    }
+    
     // 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,
+                                         !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
+                                         : &smpi_comm_null_copy_buffer_callback,
                                          request, -1.0);
+        XBT_DEBUG("recv simcall posted");
 
-    //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 %f ", request->size, smpi_or(request->size));
-    }
 
   } else {
 
 
-    int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
+    int receiver = request->dst;
 
     #ifdef HAVE_TRACING
-      int rank = smpi_process_index();
+      int rank = request->src;
       if (TRACE_smpi_view_internals()) {
         TRACE_smpi_send(rank, rank, receiver,request->size);
       }
     #endif
-/*    if(receiver == MPI_UNDEFINED) {*/
-/*      XBT_WARN("Trying to send a message to a wrong rank");*/
-/*      return;*/
-/*    }*/
     print_request("New send", request);
-    if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
-      mailbox = smpi_process_remote_mailbox_small(receiver);
+    
+        //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 %f ", request->size, smpi_os(request->size));
+    }
+    
+    if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
+      mailbox = smpi_process_remote_mailbox(receiver);
+      XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
+      smx_action_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
+      if(action ==NULL){
+       if (! (request->flags & SSEND)){
+         mailbox = smpi_process_remote_mailbox_small(receiver);
+         XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
+       } else{
+         mailbox = smpi_process_remote_mailbox_small(receiver);
+         XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
+         action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
+         if(action ==NULL){
+           XBT_DEBUG("No, we are first, send to large mailbox");
+           mailbox = smpi_process_remote_mailbox(receiver);
+         }
+       }
+      }else{
+        XBT_DEBUG("Yes there was something for us in the large mailbox");
+      }
     }else{
-      XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
       mailbox = smpi_process_remote_mailbox(receiver);
+      XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
     }
 
     void* buf = request->buf;
     if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
       void *oldbuf = NULL;
       request->detached = 1;
+      XBT_DEBUG("Send request %p is detached", request);
       request->refcount++;
       if(request->old_type->has_subtype == 0){
         oldbuf = request->buf;
-        if (!_xbt_replay_is_active() && oldbuf && request->size!=0){
+        if (!smpi_process_get_replaying() && oldbuf && request->size!=0){
+          if((smpi_privatize_global_variables)
+      	      && ((char*)request->buf >= start_data_exe)
+	      && ((char*)request->buf < start_data_exe + size_data_exe )){
+            XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
+       	    smpi_switch_data_segment(request->src);
+  	  }
           buf = xbt_malloc(request->size);
           memcpy(buf,oldbuf,request->size);
+          XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
         }
       }
-      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 %f ", request->size, smpi_os(request->size));
-    }
-
     request->action =
-      simcall_comm_isend(mailbox, request->size, -1.0,
+      simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
                          buf, request->real_size,
                          &match_send,
-                         &xbt_free, // how to free the userdata if a detached send fails
+                         &xbt_free_f, // how to free the userdata if a detached send fails
+                         !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
+                         : &smpi_comm_null_copy_buffer_callback,
                          request,
                          // detach if msg size < eager/rdv switch limit
                          request->detached);
+    XBT_DEBUG("send simcall posted");
+
+
 
 #ifdef HAVE_TRACING
     /* FIXME: detached sends are not traceable (request->action == NULL) */
@@ -448,6 +525,38 @@ void smpi_mpi_request_free(MPI_Request * request)
   }
 }
 
+
+MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype,
+                            int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
+{
+  MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
+  if(op==MPI_OP_NULL){
+    request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, src, dst, tag,
+                            comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
+  }else{
+    request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype,  src, dst, tag,
+                            comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
+    request->op = op;
+  }
+  return request;
+}
+
+MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype,
+                            int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
+{
+  MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
+  if(op==MPI_OP_NULL){
+    request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype,  src, dst, tag,
+                            comm, RMA | NON_PERSISTENT | RECV | PREPARED);
+  }else{
+    request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype,  src, dst, tag,
+                            comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
+    request->op = op;
+  }
+  return request;
+}
+
+
 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
                             int dst, int tag, MPI_Comm comm)
 {
@@ -477,8 +586,6 @@ MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
   return request;
 }
 
-
-
 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
                             int src, int tag, MPI_Comm comm)
 {
@@ -586,14 +693,28 @@ static void finish_wait(MPI_Request * request, MPI_Status * status)
     print_request("Finishing", req);
     MPI_Datatype datatype = req->old_type;
 
-    if(datatype->has_subtype == 1){
-      // This part handles the problem of non-contignous memory
-      // the unserialization at the reception
-      s_smpi_subtype_t *subtype = datatype->substruct;
-      if(req->flags & RECV) {
-        subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
+    if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){
+      if (!smpi_process_get_replaying()){
+        if( smpi_privatize_global_variables
+            && ((char*)req->old_buf >= start_data_exe)
+            && ((char*)req->old_buf < start_data_exe + size_data_exe )
+        ){
+            XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
+            smpi_switch_data_segment(smpi_process_index());
+        }
+      }
+
+      if(datatype->has_subtype == 1){
+        // This part handles the problem of non-contignous memory
+        // the unserialization at the reception
+        s_smpi_subtype_t *subtype = datatype->substruct;
+        if(req->flags & RECV)
+          subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct, req->op);
+        if(req->detached == 0) free(req->buf);
+      }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
+          int n =req->real_size/smpi_datatype_size(datatype);
+          smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
       }
-      if(req->detached == 0) free(req->buf);
     }
     smpi_comm_unuse(req->comm);
     smpi_datatype_unuse(datatype);
@@ -625,6 +746,12 @@ int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
   int flag;
 
   //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
+
+  //to avoid deadlocks
+  //multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
+  static int nsleeps = 1;
+  if(smpi_test_sleep > 0)  simcall_process_sleep(nsleeps*smpi_test_sleep);
+
   smpi_empty_status(status);
   flag = 1;
   if (!((*request)->flags & PREPARED)) {
@@ -632,8 +759,11 @@ int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
       flag = simcall_comm_test((*request)->action);
     if (flag) {
       finish_wait(request, status);
+      nsleeps=1;//reset the number of sleeps we will do next time
       if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
       *request = MPI_REQUEST_NULL;
+    }else{
+      nsleeps++;
     }
   }
   return flag;
@@ -660,6 +790,10 @@ int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
     }
   }
   if(size > 0) {
+    //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
+    static int nsleeps = 1;
+    if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
+
     i = simcall_comm_testany(comms);
     // not MPI_UNDEFINED, as this is a simix return code
     if(i != -1) {
@@ -668,6 +802,9 @@ int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
       if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
       requests[*index] = MPI_REQUEST_NULL;
       flag = 1;
+      nsleeps=1;
+    }else{
+      nsleeps++;
     }
   }else{
       //all requests are null or inactive, return true
@@ -721,12 +858,9 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
             comm, 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);
-
+  if(smpi_iprobe_sleep > 0)  simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
   // behave like a receive, but don't do it
   smx_rdv_t mailbox;
 
@@ -735,12 +869,12 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
     if (sg_cfg_get_int("smpi/async_small_thres")>0){
         mailbox = smpi_process_mailbox_small();
         XBT_DEBUG("trying to probe the perm recv mailbox");
-        request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
+        request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, (void*)request);
     }
     if (request->action==NULL){
 	mailbox = smpi_process_mailbox();
         XBT_DEBUG("trying to probe the other mailbox");
-        request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
+        request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
     }
 
   if(request->action){
@@ -1006,12 +1140,13 @@ void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
       displs[i] = count;
       count += recvcounts[i];
     }
-    tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
+    tmpbuf=(void*)smpi_get_tmp_sendbuffer(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);
+    smpi_free_tmp_buffer(tmpbuf);
 }
 
 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
@@ -1241,7 +1376,7 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
 
   char* sendtmpbuf = (char*) sendbuf;
   if( sendbuf == MPI_IN_PLACE ) {
-    sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
+    sendtmpbuf = (char *)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype));
     smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
   }
 
@@ -1272,7 +1407,10 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
       if(src != root) {
         // FIXME: possibly overkill we we have contiguous/noncontiguous data
         //  mapping...
-        tmpbufs[index] = xbt_malloc(count * dataext);
+   	    if (!smpi_process_get_replaying())
+          tmpbufs[index] = xbt_malloc(count * dataext);
+   	    else
+   	      tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
         requests[index] =
           smpi_irecv_init(tmpbufs[index], count, datatype, src,
                           system_tag, comm);
@@ -1292,14 +1430,14 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
       if(op) /* op can be MPI_OP_NULL that does nothing */
         smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
     }
-    for(index = 0; index < size - 1; index++) {
-      xbt_free(tmpbufs[index]);
-    }
+      for(index = 0; index < size - 1; index++) {
+        smpi_free_tmp_buffer(tmpbufs[index]);
+      }
     xbt_free(tmpbufs);
     xbt_free(requests);
 
     if( sendbuf == MPI_IN_PLACE ) {
-      xbt_free(sendtmpbuf);
+      smpi_free_tmp_buffer(sendtmpbuf);
     }
   }
 }
@@ -1336,7 +1474,7 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
   for(other = 0; other < rank; other++) {
     // FIXME: possibly overkill we we have contiguous/noncontiguous data
     // mapping...
-    tmpbufs[index] = xbt_malloc(count * dataext);
+    tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
     requests[index] =
       smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
                       comm);
@@ -1371,7 +1509,7 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
     }
   }
   for(index = 0; index < rank; index++) {
-    xbt_free(tmpbufs[index]);
+    smpi_free_tmp_buffer(tmpbufs[index]);
   }
   for(index = 0; index < size-1; index++) {
     smpi_mpi_request_free(&requests[index]);
@@ -1402,7 +1540,7 @@ void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
   for(other = 0; other < rank; other++) {
     // FIXME: possibly overkill we we have contiguous/noncontiguous data
     // mapping...
-    tmpbufs[index] = xbt_malloc(count * dataext);
+    tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
     requests[index] =
       smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
                       comm);
@@ -1443,7 +1581,7 @@ void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
     }
   }
   for(index = 0; index < rank; index++) {
-    xbt_free(tmpbufs[index]);
+    smpi_free_tmp_buffer(tmpbufs[index]);
   }
   for(index = 0; index < size-1; index++) {
     smpi_mpi_request_free(&requests[index]);
@@ -1451,3 +1589,62 @@ void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
   xbt_free(tmpbufs);
   xbt_free(requests);
 }
+
+
+int smpi_keyval_create(MPI_Copy_function* copy_fn, MPI_Delete_function* delete_fn, int* keyval, void* extra_state){
+
+  if(!smpi_keyvals)
+  smpi_keyvals = xbt_dict_new();
+  
+  smpi_key_elem value = (smpi_key_elem) xbt_new0(s_smpi_mpi_key_elem_t,1);
+  
+  value->copy_fn=copy_fn;
+  value->delete_fn=delete_fn;
+  
+  *keyval = keyval_id;
+  
+  xbt_dict_set(smpi_keyvals,(const char*)keyval,(void*)value, NULL);
+  keyval_id++;
+  return MPI_SUCCESS;
+}
+
+int smpi_keyval_free(int* keyval){
+  smpi_key_elem elem = xbt_dict_get_or_null(smpi_keyvals, (const char*)keyval);
+  if(!elem)
+    return MPI_ERR_ARG;
+  xbt_dict_remove(smpi_keyvals, (const char*)keyval);
+  xbt_free(elem);
+  return MPI_SUCCESS;
+}
+
+int smpi_attr_delete(MPI_Comm comm, int keyval){
+  smpi_key_elem elem = xbt_dict_get_or_null(smpi_keyvals, (const char*)&keyval);
+  if(!elem)
+    return MPI_ERR_ARG;
+  if(elem->delete_fn!=MPI_NULL_DELETE_FN){
+    void * value;
+    int flag;
+    if(smpi_attr_get(comm, keyval, &value, &flag)==MPI_SUCCESS)
+      elem->delete_fn(comm, keyval, &value, &flag);
+  }  
+  return smpi_comm_attr_delete(comm, keyval);;
+}
+
+int smpi_attr_get(MPI_Comm comm, int keyval, void* attr_value, int* flag){
+  smpi_key_elem elem = xbt_dict_get_or_null(smpi_keyvals, (const char*)&keyval);
+  if(!elem)
+    return MPI_ERR_ARG;
+  return smpi_comm_attr_get(comm, keyval, attr_value, flag);;
+}
+
+int smpi_attr_put(MPI_Comm comm, int keyval, void* attr_value){
+
+  if(!smpi_keyvals)
+  smpi_keyvals = xbt_dict_new();
+  
+  smpi_key_elem elem = xbt_dict_get_or_null(smpi_keyvals, (const char*)&keyval);
+  if(!elem )
+    return MPI_ERR_ARG;
+  return smpi_comm_attr_put(comm, keyval, attr_value);;
+}
+