diff --git a/ompi/mca/coll/libnbc/nbc_iallgather.c b/ompi/mca/coll/libnbc/nbc_iallgather.c
index 999f12d316..4b9f2224a0 100644
--- a/ompi/mca/coll/libnbc/nbc_iallgather.c
+++ b/ompi/mca/coll/libnbc/nbc_iallgather.c
@@ -70,7 +70,7 @@ static int nbc_iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendt
   if (inplace) {
     sendtype = recvtype;
     sendcount = recvcount;
-  } else {
+  } else if (!persistent) { /* for persistent, the copy must be scheduled */
     /* copy my data to receive buffer */
     rbuf = (char *) recvbuf + rank * recvcount * rcvext;
     res = NBC_Copy (sendbuf, sendcount, sendtype, rbuf, recvcount, recvtype, comm);
@@ -78,7 +78,7 @@ static int nbc_iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendt
       return res;
     }
   }
-  if (1 == p) {
+  if (1 == p && (!persistent || inplace)) {
     return nbc_get_noop_request(persistent, request);
   }
 
@@ -99,6 +99,17 @@ static int nbc_iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendt
     }
 
     sbuf = (char *)recvbuf + rank * recvcount * rcvext;
+
+    if (persistent && !inplace) { /* for nonblocking, data has been copied already */
+      /* copy my data to receive buffer (= send buffer of NBC_Sched_send) */
+      res = NBC_Sched_copy ((void *)sendbuf, false, sendcount, sendtype,
+                            sbuf, false, recvcount, recvtype, schedule, true);
+      if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+        OBJ_RELEASE(schedule);
+        return res;
+      }
+    }
+
     /* do p-1 rounds */
     for(int r = 0 ; r < p ; ++r) {
       if(r != rank) {
diff --git a/ompi/mca/coll/libnbc/nbc_iallgatherv.c b/ompi/mca/coll/libnbc/nbc_iallgatherv.c
index 42f7bc4a18..dbfc4328a0 100644
--- a/ompi/mca/coll/libnbc/nbc_iallgatherv.c
+++ b/ompi/mca/coll/libnbc/nbc_iallgatherv.c
@@ -58,7 +58,7 @@ static int nbc_iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype send
   if (inplace) {
       sendtype = recvtype;
       sendcount = recvcounts[rank];
-  } else {
+  } else if (!persistent) { /* for persistent, the copy must be scheduled */
     /* copy my data to receive buffer */
     rbuf = (char *) recvbuf + displs[rank] * rcvext;
     res = NBC_Copy (sendbuf, sendcount, sendtype, rbuf, recvcounts[rank], recvtype, comm);
@@ -74,6 +74,16 @@ static int nbc_iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype send
 
   sbuf = (char *) recvbuf + displs[rank] * rcvext;
 
+  if (persistent && !inplace) { /* for nonblocking, data has been copied already */
+    /* copy my data to receive buffer (= send buffer of NBC_Sched_send) */
+    res = NBC_Sched_copy ((void *)sendbuf, false, sendcount, sendtype,
+                          sbuf, false, recvcounts[rank], recvtype, schedule, true);
+    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+      OBJ_RELEASE(schedule);
+      return res;
+    }
+  }
+
   /* do p-1 rounds */
   for (int r = 1 ; r < p ; ++r) {
     speer = (rank + r) % p;
diff --git a/ompi/mca/coll/libnbc/nbc_iallreduce.c b/ompi/mca/coll/libnbc/nbc_iallreduce.c
index 1f4fe5d71e..b37eb4ee17 100644
--- a/ompi/mca/coll/libnbc/nbc_iallreduce.c
+++ b/ompi/mca/coll/libnbc/nbc_iallreduce.c
@@ -87,7 +87,7 @@ static int nbc_iallreduce(const void* sendbuf, void* recvbuf, int count, MPI_Dat
     return res;
   }
 
-  if (1 == p) {
+  if (1 == p && (!persistent || inplace)) {
     if (!inplace) {
       /* for a single node - copy data to receivebuf */
       res = NBC_Copy(sendbuf, count, datatype, recvbuf, count, datatype, comm);
@@ -127,13 +127,18 @@ static int nbc_iallreduce(const void* sendbuf, void* recvbuf, int count, MPI_Dat
       return OMPI_ERR_OUT_OF_RESOURCE;
     }
 
-    switch(alg) {
-      case NBC_ARED_BINOMIAL:
-        res = allred_sched_diss(rank, p, count, datatype, gap, sendbuf, recvbuf, op, inplace, schedule, tmpbuf);
-        break;
-      case NBC_ARED_RING:
-        res = allred_sched_ring(rank, p, count, datatype, sendbuf, recvbuf, op, size, ext, schedule, tmpbuf);
-        break;
+    if (p == 1) {
+      res = NBC_Sched_copy((void *)sendbuf, false, count, datatype,
+                           recvbuf, false, count, datatype, schedule, false);
+    } else {
+      switch(alg) {
+        case NBC_ARED_BINOMIAL:
+          res = allred_sched_diss(rank, p, count, datatype, gap, sendbuf, recvbuf, op, inplace, schedule, tmpbuf);
+          break;
+        case NBC_ARED_RING:
+          res = allred_sched_ring(rank, p, count, datatype, sendbuf, recvbuf, op, size, ext, schedule, tmpbuf);
+          break;
+      }
     }
 
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -348,7 +353,9 @@ static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype dat
     rbuf = recvbuf;
     tmprbuf = false;
     if (inplace) {
-        res = NBC_Copy(rbuf, count, datatype, ((char *)tmpbuf) - gap, count, datatype, MPI_COMM_SELF);
+        res = NBC_Sched_copy(rbuf, false, count, datatype,
+                             ((char *)tmpbuf) - gap, false, count, datatype,
+                             schedule, true);
         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
           return res;
         }
diff --git a/ompi/mca/coll/libnbc/nbc_ialltoall.c b/ompi/mca/coll/libnbc/nbc_ialltoall.c
index 120bf223ee..834ce51c09 100644
--- a/ompi/mca/coll/libnbc/nbc_ialltoall.c
+++ b/ompi/mca/coll/libnbc/nbc_ialltoall.c
@@ -110,16 +110,6 @@ static int nbc_ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendty
   } else
     alg = NBC_A2A_LINEAR; /*NBC_A2A_PAIRWISE;*/
 
-  if (!inplace) {
-    /* copy my data to receive buffer */
-    rbuf = (char *) recvbuf + rank * recvcount * rcvext;
-    sbuf = (char *) sendbuf + rank * sendcount * sndext;
-    res = NBC_Copy (sbuf, sendcount, sendtype, rbuf, recvcount, recvtype, comm);
-    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-      return res;
-    }
-  }
-
   /* allocate temp buffer if we need one */
   if (alg == NBC_A2A_INPLACE) {
     span = opal_datatype_span(&recvtype->super, recvcount, &gap);
@@ -205,6 +195,19 @@ static int nbc_ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendty
       return OMPI_ERR_OUT_OF_RESOURCE;
     }
 
+    if (!inplace) {
+      /* copy my data to receive buffer */
+      rbuf = (char *) recvbuf + rank * recvcount * rcvext;
+      sbuf = (char *) sendbuf + rank * sendcount * sndext;
+      res = NBC_Sched_copy (sbuf, false, sendcount, sendtype,
+                            rbuf, false, recvcount, recvtype, schedule, false);
+      if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+        OBJ_RELEASE(schedule);
+        free(tmpbuf);
+        return res;
+      }
+    }
+
     switch(alg) {
       case NBC_A2A_INPLACE:
         res = a2a_sched_inplace(rank, p, schedule, recvbuf, recvcount, recvtype, rcvext, gap, comm);
diff --git a/ompi/mca/coll/libnbc/nbc_ialltoallv.c b/ompi/mca/coll/libnbc/nbc_ialltoallv.c
index a27a8a9982..ce036dafe0 100644
--- a/ompi/mca/coll/libnbc/nbc_ialltoallv.c
+++ b/ompi/mca/coll/libnbc/nbc_ialltoallv.c
@@ -89,14 +89,6 @@ static int nbc_ialltoallv(const void* sendbuf, const int *sendcounts, const int
       NBC_Error("MPI Error in ompi_datatype_type_extent() (%i)", res);
       return res;
     }
-    if (sendcounts[rank] != 0) {
-      rbuf = (char *) recvbuf + rdispls[rank] * rcvext;
-      sbuf = (char *) sendbuf + sdispls[rank] * sndext;
-      res = NBC_Copy (sbuf, sendcounts[rank], sendtype, rbuf, recvcounts[rank], recvtype, comm);
-      if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-        return res;
-      }
-    }
   }
 
   schedule = OBJ_NEW(NBC_Schedule);
@@ -106,6 +98,17 @@ static int nbc_ialltoallv(const void* sendbuf, const int *sendcounts, const int
   }
 
 
+  if (!inplace && sendcounts[rank] != 0) {
+    rbuf = (char *) recvbuf + rdispls[rank] * rcvext;
+    sbuf = (char *) sendbuf + sdispls[rank] * sndext;
+    res = NBC_Sched_copy (sbuf, false, sendcounts[rank], sendtype,
+                          rbuf, false, recvcounts[rank], recvtype, schedule, false);
+    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+      OBJ_RELEASE(schedule);
+      return res;
+    }
+  }
+
   if (inplace) {
     res = a2av_sched_inplace(rank, p, schedule, recvbuf, recvcounts,
                                  rdispls, rcvext, recvtype, gap);
diff --git a/ompi/mca/coll/libnbc/nbc_ialltoallw.c b/ompi/mca/coll/libnbc/nbc_ialltoallw.c
index 6ce70e9495..de94206c41 100644
--- a/ompi/mca/coll/libnbc/nbc_ialltoallw.c
+++ b/ompi/mca/coll/libnbc/nbc_ialltoallw.c
@@ -77,13 +77,6 @@ static int nbc_ialltoallw(const void* sendbuf, const int *sendcounts, const int
     sendcounts = recvcounts;
     sdispls = rdispls;
     sendtypes = recvtypes;
-  } else if (sendcounts[rank] != 0) {
-    rbuf = (char *) recvbuf + rdispls[rank];
-    sbuf = (char *) sendbuf + sdispls[rank];
-    res = NBC_Copy(sbuf, sendcounts[rank], sendtypes[rank], rbuf, recvcounts[rank], recvtypes[rank], comm);
-    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-      return res;
-    }
   }
 
   schedule = OBJ_NEW(NBC_Schedule);
@@ -92,6 +85,16 @@ static int nbc_ialltoallw(const void* sendbuf, const int *sendcounts, const int
     return OMPI_ERR_OUT_OF_RESOURCE;
   }
 
+  if (!inplace && sendcounts[rank] != 0) {
+    rbuf = (char *) recvbuf + rdispls[rank];
+    sbuf = (char *) sendbuf + sdispls[rank];
+    res = NBC_Sched_copy(sbuf, false, sendcounts[rank], sendtypes[rank],
+                         rbuf, false, recvcounts[rank], recvtypes[rank], schedule, false);
+    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+      return res;
+    }
+  }
+
   if (inplace) {
     res = a2aw_sched_inplace(rank, p, schedule, recvbuf,
                                  recvcounts, rdispls, recvtypes);
diff --git a/ompi/mca/coll/libnbc/nbc_iexscan.c b/ompi/mca/coll/libnbc/nbc_iexscan.c
index 6313f8a161..faad1e6a6b 100644
--- a/ompi/mca/coll/libnbc/nbc_iexscan.c
+++ b/ompi/mca/coll/libnbc/nbc_iexscan.c
@@ -64,23 +64,6 @@ static int nbc_iexscan(const void* sendbuf, void* recvbuf, int count, MPI_Dataty
     rank = ompi_comm_rank (comm);
     p = ompi_comm_size (comm);
 
-    span = opal_datatype_span(&datatype->super, count, &gap);
-    if (0 < rank) {
-        tmpbuf = malloc(span);
-        if (NULL == tmpbuf) {
-            return OMPI_ERR_OUT_OF_RESOURCE;
-        }
-        if (inplace) {
-            res = NBC_Copy(recvbuf, count, datatype, (char *)tmpbuf-gap, count, datatype, comm);
-        } else {
-            res = NBC_Copy(sendbuf, count, datatype, (char *)tmpbuf-gap, count, datatype, comm);
-        }
-        if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-            free(tmpbuf);
-            return res;
-        }
-    }
-
 #ifdef NBC_CACHE_SCHEDULE
     /* search schedule in communicator specific tree */
     search.sendbuf = sendbuf;
@@ -98,6 +81,24 @@ static int nbc_iexscan(const void* sendbuf, void* recvbuf, int count, MPI_Dataty
         }
 
         if (rank != 0) {
+            span = opal_datatype_span(&datatype->super, count, &gap);
+            tmpbuf = malloc(span);
+            if (NULL == tmpbuf) {
+                return OMPI_ERR_OUT_OF_RESOURCE;
+            }
+            if (inplace) {
+                res = NBC_Sched_copy(recvbuf, false, count, datatype,
+                                     (char *)tmpbuf-gap, false, count, datatype, schedule, false);
+            } else {
+                res = NBC_Sched_copy((void *)sendbuf, false, count, datatype,
+                                     (char *)tmpbuf-gap, false, count, datatype, schedule, false);
+            }
+            if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+                OBJ_RELEASE(schedule);
+                free(tmpbuf);
+                return res;
+            }
+
             res = NBC_Sched_recv (recvbuf, false, count, datatype, rank-1, schedule, false);
 
             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
diff --git a/ompi/mca/coll/libnbc/nbc_igather.c b/ompi/mca/coll/libnbc/nbc_igather.c
index 13802e0b8b..107c1a1f70 100644
--- a/ompi/mca/coll/libnbc/nbc_igather.c
+++ b/ompi/mca/coll/libnbc/nbc_igather.c
@@ -71,13 +71,6 @@ static int nbc_igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype
   if (inplace) {
     sendcount = recvcount;
     sendtype = recvtype;
-  } else if (rank == root) {
-    rbuf = ((char *)recvbuf) + (rank*recvcount*rcvext);
-    /* if I am the root - just copy the message (only without MPI_IN_PLACE) */
-    res = NBC_Copy(sendbuf, sendcount, sendtype, rbuf, recvcount, recvtype, comm);
-    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-      return res;
-    }
   }
 
 #ifdef NBC_CACHE_SCHEDULE
@@ -111,7 +104,17 @@ static int nbc_igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype
     } else {
       for (int i = 0 ; i < p ; ++i) {
         rbuf = (char *)recvbuf + i * recvcount * rcvext;
-        if (i != root) {
+        if (i == root) {
+          if (!inplace) {
+            /* if I am the root - just copy the message */
+            res = NBC_Sched_copy ((void *)sendbuf, false, sendcount, sendtype,
+                                  rbuf, false, recvcount, recvtype, schedule, false);
+            if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+              OBJ_RELEASE(schedule);
+              return res;
+            }
+          }
+        } else {
           /* root receives message to the right buffer */
           res = NBC_Sched_recv (rbuf, false, recvcount, recvtype, i, schedule, false);
           if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
diff --git a/ompi/mca/coll/libnbc/nbc_igatherv.c b/ompi/mca/coll/libnbc/nbc_igatherv.c
index bc3880ab9f..a1b916429f 100644
--- a/ompi/mca/coll/libnbc/nbc_igatherv.c
+++ b/ompi/mca/coll/libnbc/nbc_igatherv.c
@@ -72,8 +72,8 @@ static int nbc_igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtyp
       if (i == root) {
         if (!inplace) {
           /* if I am the root - just copy the message */
-          res = NBC_Copy (sendbuf, sendcount, sendtype, rbuf, recvcounts[i], recvtype,
-                          comm);
+          res = NBC_Sched_copy ((void *)sendbuf, false, sendcount, sendtype,
+                                rbuf, false, recvcounts[i], recvtype, schedule, false);
           if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
             OBJ_RELEASE(schedule);
             return res;
diff --git a/ompi/mca/coll/libnbc/nbc_ireduce.c b/ompi/mca/coll/libnbc/nbc_ireduce.c
index 7f74a37125..042aacba3f 100644
--- a/ompi/mca/coll/libnbc/nbc_ireduce.c
+++ b/ompi/mca/coll/libnbc/nbc_ireduce.c
@@ -85,7 +85,7 @@ static int nbc_ireduce(const void* sendbuf, void* recvbuf, int count, MPI_Dataty
   }
 
   /* only one node -> copy data */
-  if (p == 1) {
+  if (1 == p && (!persistent || inplace)) {
     if (!inplace) {
       res = NBC_Copy (sendbuf, count, datatype, recvbuf, count, datatype, comm);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -140,13 +140,18 @@ static int nbc_ireduce(const void* sendbuf, void* recvbuf, int count, MPI_Dataty
       return OMPI_ERR_OUT_OF_RESOURCE;
     }
 
-    switch(alg) {
-      case NBC_RED_BINOMIAL:
-        res = red_sched_binomial(rank, p, root, sendbuf, redbuf, tmpredbuf, count, datatype, op, inplace, schedule, tmpbuf);
-        break;
-      case NBC_RED_CHAIN:
-        res = red_sched_chain(rank, p, root, sendbuf, recvbuf, count, datatype, op, ext, size, schedule, tmpbuf, segsize);
-        break;
+    if (p == 1) {
+      res = NBC_Sched_copy ((void *)sendbuf, false, count, datatype,
+                            recvbuf, false, count, datatype, schedule, false);
+    } else {
+      switch(alg) {
+        case NBC_RED_BINOMIAL:
+          res = red_sched_binomial(rank, p, root, sendbuf, redbuf, tmpredbuf, count, datatype, op, inplace, schedule, tmpbuf);
+          break;
+        case NBC_RED_CHAIN:
+          res = red_sched_chain(rank, p, root, sendbuf, recvbuf, count, datatype, op, ext, size, schedule, tmpbuf, segsize);
+          break;
+      }
     }
 
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -351,7 +356,9 @@ static inline int red_sched_binomial (int rank, int p, int root, const void *sen
     rbuf = redbuf;
     tmprbuf = tmpredbuf;
     if (inplace) {
-        res = NBC_Copy(rbuf, count, datatype, ((char *)tmpbuf)-gap, count, datatype, MPI_COMM_SELF);
+        res = NBC_Sched_copy(rbuf, false, count, datatype,
+                             ((char *)tmpbuf)-gap, false, count, datatype,
+                             schedule, true);
         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
           return res;
         }
diff --git a/ompi/mca/coll/libnbc/nbc_ireduce_scatter.c b/ompi/mca/coll/libnbc/nbc_ireduce_scatter.c
index b9025c2493..f22e198477 100644
--- a/ompi/mca/coll/libnbc/nbc_ireduce_scatter.c
+++ b/ompi/mca/coll/libnbc/nbc_ireduce_scatter.c
@@ -70,7 +70,7 @@ static int nbc_ireduce_scatter(const void* sendbuf, void* recvbuf, const int *re
     count += recvcounts[r];
   }
 
-  if (p == 1 || 0 == count) {
+  if ((1 == p && (!persistent || inplace)) || 0 == count) {
     if (!inplace) {
       /* single node not in_place: copy data to recvbuf */
       res = NBC_Copy(sendbuf, recvcounts[0], datatype, recvbuf, recvcounts[0], datatype, comm);
@@ -174,8 +174,14 @@ static int nbc_ireduce_scatter(const void* sendbuf, void* recvbuf, const int *re
       }
     }
 
-    res = NBC_Sched_copy (lbuf, true, recvcounts[0], datatype, recvbuf, false,
-                          recvcounts[0], datatype, schedule, false);
+    if (p == 1) {
+      /* single node not in_place: copy data to recvbuf */
+      res = NBC_Sched_copy ((void *)sendbuf, false, recvcounts[0], datatype,
+                            recvbuf, false, recvcounts[0], datatype, schedule, false);
+    } else {
+      res = NBC_Sched_copy (lbuf, true, recvcounts[0], datatype, recvbuf, false,
+                            recvcounts[0], datatype, schedule, false);
+    }
   } else {
     res = NBC_Sched_recv (recvbuf, false, recvcounts[rank], datatype, 0, schedule, false);
   }
diff --git a/ompi/mca/coll/libnbc/nbc_ireduce_scatter_block.c b/ompi/mca/coll/libnbc/nbc_ireduce_scatter_block.c
index cde6bd914e..e40cc69228 100644
--- a/ompi/mca/coll/libnbc/nbc_ireduce_scatter_block.c
+++ b/ompi/mca/coll/libnbc/nbc_ireduce_scatter_block.c
@@ -89,7 +89,8 @@ static int nbc_ireduce_scatter_block(const void* sendbuf, void* recvbuf, int rec
 
     /* copy data to redbuf if we only have a single node */
     if ((p == 1) && !inplace) {
-      res = NBC_Copy (sendbuf, count, datatype, redbuf, count, datatype, comm);
+      res = NBC_Sched_copy ((void *)sendbuf, false, count, datatype,
+                            redbuf, false, count, datatype, schedule, false);
       if (OMPI_SUCCESS != res) {
         OBJ_RELEASE(schedule);
         free(tmpbuf);
diff --git a/ompi/mca/coll/libnbc/nbc_iscan.c b/ompi/mca/coll/libnbc/nbc_iscan.c
index a2fdd9aac6..7b4d8bf0ab 100644
--- a/ompi/mca/coll/libnbc/nbc_iscan.c
+++ b/ompi/mca/coll/libnbc/nbc_iscan.c
@@ -61,14 +61,6 @@ static int nbc_iscan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype
   rank = ompi_comm_rank (comm);
   p = ompi_comm_size (comm);
 
-  if (!inplace) {
-    /* copy data to receivebuf */
-    res = NBC_Copy (sendbuf, count, datatype, recvbuf, count, datatype, comm);
-    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-      return res;
-    }
-  }
-
 #ifdef NBC_CACHE_SCHEDULE
   NBC_Scan_args *args, *found, search;
 
@@ -86,6 +78,16 @@ static int nbc_iscan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype
       return OMPI_ERR_OUT_OF_RESOURCE;
     }
 
+    if (!inplace) {
+      /* copy data to receivebuf */
+      res = NBC_Sched_copy ((void *)sendbuf, false, count, datatype,
+                            recvbuf, false, count, datatype, schedule, false);
+      if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+        OBJ_RELEASE(schedule);
+        return res;
+      }
+    }
+
     if(rank != 0) {
       span = opal_datatype_span(&datatype->super, count, &gap);
       tmpbuf = malloc (span);
diff --git a/ompi/mca/coll/libnbc/nbc_iscatter.c b/ompi/mca/coll/libnbc/nbc_iscatter.c
index dda22a3827..8c4fd138f5 100644
--- a/ompi/mca/coll/libnbc/nbc_iscatter.c
+++ b/ompi/mca/coll/libnbc/nbc_iscatter.c
@@ -70,15 +70,6 @@ static int nbc_iscatter (const void* sendbuf, int sendcount, MPI_Datatype sendty
     }
   }
 
-  if ((rank == root) && (!inplace)) {
-    sbuf = (char *) sendbuf + rank * sendcount * sndext;
-    /* if I am the root - just copy the message (not for MPI_IN_PLACE) */
-    res = NBC_Copy (sbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
-    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-      return res;
-    }
-  }
-
 #ifdef NBC_CACHE_SCHEDULE
   NBC_Scatter_args *args, *found, search;
 
@@ -109,7 +100,17 @@ static int nbc_iscatter (const void* sendbuf, int sendcount, MPI_Datatype sendty
     } else {
       for (int i = 0 ; i < p ; ++i) {
         sbuf = (char *) sendbuf + i * sendcount * sndext;
-        if (i != root) {
+        if (i == root) {
+          if (!inplace) {
+            /* if I am the root - just copy the message */
+            res = NBC_Sched_copy (sbuf, false, sendcount, sendtype,
+                                  recvbuf, false, recvcount, recvtype, schedule, false);
+            if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+              OBJ_RELEASE(schedule);
+              return res;
+            }
+          }
+        } else {
           /* root sends the right buffer to the right receiver */
           res = NBC_Sched_send (sbuf, false, sendcount, sendtype, i, schedule, false);
           if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
diff --git a/ompi/mca/coll/libnbc/nbc_iscatterv.c b/ompi/mca/coll/libnbc/nbc_iscatterv.c
index 08faacb18e..1a1959ecba 100644
--- a/ompi/mca/coll/libnbc/nbc_iscatterv.c
+++ b/ompi/mca/coll/libnbc/nbc_iscatterv.c
@@ -64,7 +64,8 @@ static int nbc_iscatterv(const void* sendbuf, const int *sendcounts, const int *
       if (i == root) {
         if (!inplace) {
           /* if I am the root - just copy the message */
-          res = NBC_Copy (sbuf, sendcounts[i], sendtype, recvbuf, recvcount, recvtype, comm);
+          res = NBC_Sched_copy (sbuf, false, sendcounts[i], sendtype,
+                                recvbuf, false, recvcount, recvtype, schedule, false);
         } else {
           res = OMPI_SUCCESS;
         }