diff --git a/ompi/mca/coll/libnbc/nbc.c b/ompi/mca/coll/libnbc/nbc.c
index ef9ace0ec4..6fb717b218 100644
--- a/ompi/mca/coll/libnbc/nbc.c
+++ b/ompi/mca/coll/libnbc/nbc.c
@@ -10,7 +10,7 @@
  *                         rights reserved.
  * Copyright (c) 2015      Los Alamos National Security, LLC.  All rights
  *                         reserved.
- * Copyright (c) 2015      Research Organization for Information Science
+ * Copyright (c) 2015-2016 Research Organization for Information Science
  *                         and Technology (RIST). All rights reserved.
  *
  * Author(s): Torsten Hoefler <htor@cs.indiana.edu>
@@ -111,7 +111,7 @@ static int nbc_schedule_round_append (NBC_Schedule *schedule, void *data, int da
 }
 
 /* this function puts a send into the schedule */
-int NBC_Sched_send (const void* buf, char tmpbuf, int count, MPI_Datatype datatype, int dest, NBC_Schedule *schedule, bool barrier) {
+static int NBC_Sched_send_internal (const void* buf, char tmpbuf, int count, MPI_Datatype datatype, int dest, bool local, NBC_Schedule *schedule, bool barrier) {
   NBC_Args_send send_args;
   int ret;
 
@@ -122,6 +122,7 @@ int NBC_Sched_send (const void* buf, char tmpbuf, int count, MPI_Datatype dataty
   send_args.count = count;
   send_args.datatype = datatype;
   send_args.dest = dest;
+  send_args.local = local;
 
   /* append to the round-schedule */
   ret = nbc_schedule_round_append (schedule, &send_args, sizeof (send_args), barrier);
@@ -134,8 +135,16 @@ int NBC_Sched_send (const void* buf, char tmpbuf, int count, MPI_Datatype dataty
   return OMPI_SUCCESS;
 }
 
+int NBC_Sched_send (const void* buf, char tmpbuf, int count, MPI_Datatype datatype, int dest, NBC_Schedule *schedule, bool barrier) {
+  return NBC_Sched_send_internal (buf, tmpbuf, count, datatype, dest, false, schedule, barrier);
+}
+
+int NBC_Sched_local_send (const void* buf, char tmpbuf, int count, MPI_Datatype datatype, int dest, NBC_Schedule *schedule, bool barrier) {
+  return NBC_Sched_send_internal (buf, tmpbuf, count, datatype, dest, true, schedule, barrier);
+}
+
 /* this function puts a receive into the schedule */
-int NBC_Sched_recv (void* buf, char tmpbuf, int count, MPI_Datatype datatype, int source, NBC_Schedule *schedule, bool barrier) {
+static int NBC_Sched_recv_internal (void* buf, char tmpbuf, int count, MPI_Datatype datatype, int source, bool local, NBC_Schedule *schedule, bool barrier) {
   NBC_Args_recv recv_args;
   int ret;
 
@@ -146,6 +155,7 @@ int NBC_Sched_recv (void* buf, char tmpbuf, int count, MPI_Datatype datatype, in
   recv_args.count = count;
   recv_args.datatype = datatype;
   recv_args.source = source;
+  recv_args.local = local;
 
   /* append to the round-schedule */
   ret = nbc_schedule_round_append (schedule, &recv_args, sizeof (recv_args), barrier);
@@ -158,8 +168,16 @@ int NBC_Sched_recv (void* buf, char tmpbuf, int count, MPI_Datatype datatype, in
   return OMPI_SUCCESS;
 }
 
+int NBC_Sched_recv (void* buf, char tmpbuf, int count, MPI_Datatype datatype, int source, NBC_Schedule *schedule, bool barrier) {
+  return NBC_Sched_recv_internal(buf, tmpbuf, count, datatype, source, false, schedule, barrier);
+}
+
+int NBC_Sched_local_recv (void* buf, char tmpbuf, int count, MPI_Datatype datatype, int source, NBC_Schedule *schedule, bool barrier) {
+  return NBC_Sched_recv_internal(buf, tmpbuf, count, datatype, source, true, schedule, barrier);
+}
+
 /* this function puts an operation into the schedule */
-int NBC_Sched_op (void *buf3, char tmpbuf3, const void* buf1, char tmpbuf1, void* buf2, char tmpbuf2, int count, MPI_Datatype datatype,
+int NBC_Sched_op (const void* buf1, char tmpbuf1, void* buf2, char tmpbuf2, int count, MPI_Datatype datatype,
                   MPI_Op op, NBC_Schedule *schedule, bool barrier) {
   NBC_Args_op op_args;
   int ret;
@@ -168,10 +186,8 @@ int NBC_Sched_op (void *buf3, char tmpbuf3, const void* buf1, char tmpbuf1, void
   op_args.type = OP;
   op_args.buf1 = buf1;
   op_args.buf2 = buf2;
-  op_args.buf3 = buf3;
   op_args.tmpbuf1 = tmpbuf1;
   op_args.tmpbuf2 = tmpbuf2;
-  op_args.tmpbuf3 = tmpbuf3;
   op_args.count = count;
   op_args.op = op;
   op_args.datatype = datatype;
@@ -182,7 +198,7 @@ int NBC_Sched_op (void *buf3, char tmpbuf3, const void* buf1, char tmpbuf1, void
     return ret;
   }
 
-  NBC_DEBUG(10, "added op - ends at byte %i\n", nbc_schedule_get_size (schedule));
+  NBC_DEBUG(10, "added op2 - ends at byte %i\n", nbc_schedule_get_size (schedule));
 
   return OMPI_SUCCESS;
 }
@@ -373,7 +389,7 @@ static inline int NBC_Start_round(NBC_Handle *handle) {
   NBC_Args_op         opargs;
   NBC_Args_copy     copyargs;
   NBC_Args_unpack unpackargs;
-  void *buf1,  *buf2, *buf3;
+  void *buf1,  *buf2;
 
   /* get round-schedule address */
   ptr = handle->schedule->data + handle->row_offset;
@@ -410,7 +426,7 @@ static inline int NBC_Start_round(NBC_Handle *handle) {
         handle->req_array = tmp;
 
         res = MCA_PML_CALL(isend(buf1, sendargs.count, sendargs.datatype, sendargs.dest, handle->tag,
-                                 MCA_PML_BASE_SEND_STANDARD, handle->comm,
+                                 MCA_PML_BASE_SEND_STANDARD, sendargs.local?handle->comm->c_local_comm:handle->comm,
                                  handle->req_array+handle->req_count - 1));
         if (OMPI_SUCCESS != res) {
           NBC_Error ("Error in MPI_Isend(%lu, %i, %p, %i, %i, %lu) (%i)", (unsigned long)buf1, sendargs.count,
@@ -444,7 +460,7 @@ static inline int NBC_Start_round(NBC_Handle *handle) {
 
         handle->req_array = tmp;
 
-        res = MCA_PML_CALL(irecv(buf1, recvargs.count, recvargs.datatype, recvargs.source, handle->tag, handle->comm,
+        res = MCA_PML_CALL(irecv(buf1, recvargs.count, recvargs.datatype, recvargs.source, handle->tag, recvargs.local?handle->comm->c_local_comm:handle->comm,
                                  handle->req_array+handle->req_count-1));
         if (OMPI_SUCCESS != res) {
           NBC_Error("Error in MPI_Irecv(%lu, %i, %p, %i, %i, %lu) (%i)", (unsigned long)buf1, recvargs.count,
@@ -456,10 +472,10 @@ static inline int NBC_Start_round(NBC_Handle *handle) {
 #endif
         break;
       case OP:
-        NBC_DEBUG(5, "  OP   (offset %li) ", offset);
+        NBC_DEBUG(5, "  OP2  (offset %li) ", offset);
         NBC_GET_BYTES(ptr,opargs);
-        NBC_DEBUG(5, "*buf1: %p, buf2: %p, buf3: %p, count: %i, type: %p)\n", opargs.buf1, opargs.buf2,
-                  opargs.buf3, opargs.count, opargs.datatype);
+        NBC_DEBUG(5, "*buf1: %p, buf2: %p, count: %i, type: %p)\n", opargs.buf1, opargs.buf2,
+                  opargs.count, opargs.datatype);
         /* get buffers */
         if(opargs.tmpbuf1) {
           buf1=(char*)handle->tmpbuf+(long)opargs.buf1;
@@ -471,12 +487,7 @@ static inline int NBC_Start_round(NBC_Handle *handle) {
         } else {
           buf2=opargs.buf2;
         }
-        if(opargs.tmpbuf3) {
-          buf3=(char*)handle->tmpbuf+(long)opargs.buf3;
-        } else {
-          buf3=opargs.buf3;
-        }
-        ompi_3buff_op_reduce(opargs.op, buf1, buf2, buf3, opargs.count, opargs.datatype);
+        ompi_op_reduce(opargs.op, buf1, buf2, opargs.count, opargs.datatype);
         break;
       case COPY:
         NBC_DEBUG(5, "  COPY   (offset %li) ", offset);
diff --git a/ompi/mca/coll/libnbc/nbc_iallreduce.c b/ompi/mca/coll/libnbc/nbc_iallreduce.c
index 2e1b0dd00b..0b624f90c4 100644
--- a/ompi/mca/coll/libnbc/nbc_iallreduce.c
+++ b/ompi/mca/coll/libnbc/nbc_iallreduce.c
@@ -7,7 +7,7 @@
  *                         rights reserved.
  * Copyright (c) 2013-2015 Los Alamos National Security, LLC. All rights
  *                         reserved.
- * Copyright (c) 2014-2015 Research Organization for Information Science
+ * Copyright (c) 2014-2016 Research Organization for Information Science
  *                         and Technology (RIST). All rights reserved.
  *
  * Author(s): Torsten Hoefler <htor@cs.indiana.edu>
@@ -16,16 +16,17 @@
 #include "nbc_internal.h"
 #include "ompi/communicator/communicator.h"
 #include "ompi/datatype/ompi_datatype.h"
+#include "ompi/op/op.h"
 
 #include <assert.h>
 
-static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype datatype, const void *sendbuf,
-                                    void *recvbuf, MPI_Op op, NBC_Schedule *schedule, NBC_Handle *handle);
+static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype datatype, ptrdiff_t gap, const void *sendbuf,
+                                    void *recvbuf, MPI_Op op, char inplace, NBC_Schedule *schedule, NBC_Handle *handle);
 static inline int allred_sched_ring(int rank, int p, int count, MPI_Datatype datatype, const void *sendbuf,
                                     void *recvbuf, MPI_Op op, int size, int ext, NBC_Schedule *schedule,
                                     NBC_Handle *handle);
 static inline int allred_sched_linear(int rank, int p, const void *sendbuf, void *recvbuf, int count,
-                                      MPI_Datatype datatype, MPI_Op op, int ext, int size,
+                                      MPI_Datatype datatype, ptrdiff_t gap, MPI_Op op, int ext, int size,
                                       NBC_Schedule *schedule, NBC_Handle *handle);
 
 #ifdef NBC_CACHE_SCHEDULE
@@ -62,6 +63,7 @@ int ompi_coll_libnbc_iallreduce(const void* sendbuf, void* recvbuf, int count, M
   char inplace;
   NBC_Handle *handle;
   ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
+  ptrdiff_t span, gap;
 
   NBC_IN_PLACE(sendbuf, recvbuf, inplace);
 
@@ -85,7 +87,8 @@ int ompi_coll_libnbc_iallreduce(const void* sendbuf, void* recvbuf, int count, M
     return res;
   }
 
-  handle->tmpbuf = malloc (ext * count);
+  span = opal_datatype_span(&datatype->super, count, &gap);
+  handle->tmpbuf = malloc (span);
   if (OPAL_UNLIKELY(NULL == handle->tmpbuf)) {
     NBC_Return_handle (handle);
     return OMPI_ERR_OUT_OF_RESOURCE;
@@ -101,7 +104,7 @@ int ompi_coll_libnbc_iallreduce(const void* sendbuf, void* recvbuf, int count, M
   }
 
   /* algorithm selection */
-  if(p < 4 || size*count < 65536 || inplace) {
+  if(p < 4 || size*count < 65536 || !ompi_op_is_commute(op) || inplace) {
     alg = NBC_ARED_BINOMIAL;
   } else {
     alg = NBC_ARED_RING;
@@ -128,7 +131,7 @@ int ompi_coll_libnbc_iallreduce(const void* sendbuf, void* recvbuf, int count, M
 
     switch(alg) {
       case NBC_ARED_BINOMIAL:
-        res = allred_sched_diss(rank, p, count, datatype, sendbuf, recvbuf, op, schedule, handle);
+        res = allred_sched_diss(rank, p, count, datatype, gap, sendbuf, recvbuf, op, inplace, schedule, handle);
         break;
       case NBC_ARED_RING:
         res = allred_sched_ring(rank, p, count, datatype, sendbuf, recvbuf, op, size, ext, schedule, handle);
@@ -199,6 +202,7 @@ int ompi_coll_libnbc_iallreduce_inter(const void* sendbuf, void* recvbuf, int co
   NBC_Schedule *schedule;
   NBC_Handle *handle;
   ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
+  ptrdiff_t span, gap;
 
   rank = ompi_comm_rank (comm);
   rsize = ompi_comm_remote_size (comm);
@@ -220,7 +224,8 @@ int ompi_coll_libnbc_iallreduce_inter(const void* sendbuf, void* recvbuf, int co
     return res;
   }
 
-  handle->tmpbuf = malloc (ext * count);
+  span = opal_datatype_span(&datatype->super, count, &gap);
+  handle->tmpbuf = malloc (span);
   if (OPAL_UNLIKELY(NULL == handle->tmpbuf)) {
     NBC_Return_handle (handle);
     return OMPI_ERR_OUT_OF_RESOURCE;
@@ -235,7 +240,7 @@ int ompi_coll_libnbc_iallreduce_inter(const void* sendbuf, void* recvbuf, int co
   /* ensure the schedule is released with the handle on error */
   handle->schedule = schedule;
 
-  res = allred_sched_linear (rank, rsize, sendbuf, recvbuf, count, datatype, op,
+  res = allred_sched_linear (rank, rsize, sendbuf, recvbuf, count, datatype, gap, op,
                              ext, size, schedule, handle);
   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
     NBC_Return_handle (handle);
@@ -296,13 +301,33 @@ int ompi_coll_libnbc_iallreduce_inter(const void* sendbuf, void* recvbuf, int co
   if (vrank == 0) rank = root; \
   if (vrank == root) rank = 0; \
 }
-static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype datatype, const void *sendbuf, void *recvbuf,
-                                    MPI_Op op, NBC_Schedule *schedule, NBC_Handle *handle) {
+static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype datatype, ptrdiff_t gap, const void *sendbuf, void *recvbuf,
+                                    MPI_Op op, char inplace, NBC_Schedule *schedule, NBC_Handle *handle) {
   int root, vrank, maxr, vpeer, peer, res;
+  char *rbuf, *lbuf, *buf;
+  int tmprbuf, tmplbuf;
 
   root = 0; /* this makes the code for ireduce and iallreduce nearly identical - could be changed to improve performance */
   RANK2VRANK(rank, vrank, root);
   maxr = (int)ceil((log((double)p)/LOG2));
+  /* ensure the result ends up in recvbuf on vrank 0 */
+  if (0 == (maxr%2)) {
+    rbuf = (void *)(-gap);
+    tmprbuf = true;
+    lbuf = recvbuf;
+    tmplbuf = false;
+  } else {
+    lbuf = (void *)(-gap);
+    tmplbuf = true;
+    rbuf = recvbuf;
+    tmprbuf = false;
+    if (inplace) {
+        res = NBC_Copy(rbuf, count, datatype, ((char *)handle->tmpbuf) - gap, count, datatype, MPI_COMM_SELF);
+        if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+          return res;
+        }
+    }
+  }
 
   for (int r = 1, firstred = 1 ; r <= maxr ; ++r) {
     if ((vrank % (1 << r)) == 0) {
@@ -311,35 +336,37 @@ static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype dat
       VRANK2RANK(peer, vpeer, root)
       if (peer < p) {
         /* we have to wait until we have the data */
-        res = NBC_Sched_recv (0, true, count, datatype, peer, schedule, true);
+        res = NBC_Sched_recv (rbuf, tmprbuf, count, datatype, peer, schedule, true);
         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
           return res;
         }
 
         /* this cannot be done until handle->tmpbuf is unused :-( so barrier after the op */
-        if (firstred && MPI_IN_PLACE != sendbuf) {
+        if (firstred && !inplace) {
           /* perform the reduce with the senbuf */
-          res = NBC_Sched_op (recvbuf, false, sendbuf, false, 0, true, count, datatype, op, schedule, true);
+          res = NBC_Sched_op (sendbuf, false, rbuf, tmprbuf, count, datatype, op, schedule, true);
           firstred = 0;
         } else {
           /* perform the reduce in my local buffer */
-          res = NBC_Sched_op (recvbuf, false, recvbuf, false, 0, true, count, datatype, op, schedule, true);
+          res = NBC_Sched_op (lbuf, tmplbuf, rbuf, tmprbuf, count, datatype, op, schedule, true);
         }
-
         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
           return res;
         }
+        /* swap left and right buffers */
+        buf = rbuf; rbuf = lbuf ; lbuf = buf;
+        tmprbuf ^= 1; tmplbuf ^= 1;
       }
     } else {
       /* we have to send this round */
       vpeer = vrank - (1 << (r - 1));
       VRANK2RANK(peer, vpeer, root)
-      if (firstred && MPI_IN_PLACE != sendbuf) {
+      if (firstred && !inplace) {
         /* we have to use the sendbuf in the first round .. */
         res = NBC_Sched_send (sendbuf, false, count, datatype, peer, schedule, false);
       } else {
-        /* and the recvbuf in all remeining rounds */
-        res = NBC_Sched_send (recvbuf, false, count, datatype, peer, schedule, false);
+        /* and the recvbuf in all remaining rounds */
+        res = NBC_Sched_send (lbuf, tmplbuf, count, datatype, peer, schedule, false);
       }
 
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -373,6 +400,7 @@ static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype dat
     }
   }
 
+  if (0 == vrank) assert(lbuf == recvbuf);
   /* now send to the right hosts */
   for (int r = 0; r < maxr; ++r) {
     if (((vrank + (1 << r) < p) && (vrank < (1 << r))) || (vrank == 0)) {
@@ -550,9 +578,8 @@ static inline int allred_sched_ring (int r, int p, int count, MPI_Datatype datat
       break;
     }
 
-    res = NBC_Sched_op ((char *) recvbuf + roffset, false, (char *) sendbuf + roffset, false,
-                        (char *) recvbuf + roffset, false, segsizes[relement], datatype, op, schedule,
-                        true);
+    res = NBC_Sched_op ((char *) sendbuf + roffset, false, (char *) recvbuf + roffset, false,
+                         segsizes[relement], datatype, op, schedule, true);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
       break;
     }
@@ -590,7 +617,7 @@ static inline int allred_sched_ring (int r, int p, int count, MPI_Datatype datat
 }
 
 static inline int allred_sched_linear(int rank, int rsize, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
-				      MPI_Op op, int ext, int size, NBC_Schedule *schedule, NBC_Handle *handle) {
+				      ptrdiff_t gap, MPI_Op op, int ext, int size, NBC_Schedule *schedule, NBC_Handle *handle) {
   int res;
 
   if (0 == count) {
@@ -603,34 +630,56 @@ static inline int allred_sched_linear(int rank, int rsize, const void *sendbuf,
     return res;
   }
 
-  res = NBC_Sched_recv (recvbuf, false, count, datatype, 0, schedule, false);
+  /* recv my data to the remote root */
+  if (0 != rank || 1 ==(rsize%2)) {
+    res = NBC_Sched_recv (recvbuf, false, count, datatype, 0, schedule, false);
+  } else {
+    res = NBC_Sched_recv ((void *)(-gap), true, count, datatype, 0, schedule, false);
+  }
   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
     return res;
   }
 
   if (0 == rank) {
-    /* wait for data from the remote root */
+    char *rbuf, *lbuf, *buf;
+    int tmprbuf, tmplbuf;
+
     res = NBC_Sched_barrier (schedule);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
       return res;
     }
 
+    /* ensure the result ends up in recvbuf */
+    if (0 == (rsize%2)) {
+      lbuf = (void *)(-gap);
+      tmplbuf = true;
+      rbuf = recvbuf;
+      tmprbuf = false;
+    } else {
+      rbuf = (void *)(-gap);
+      tmprbuf = true;
+      lbuf = recvbuf;
+      tmplbuf = false;
+    }
+
     /* get data from remote peers and reduce */
     for (int rpeer = 1 ; rpeer < rsize ; ++rpeer) {
-      res = NBC_Sched_recv (0, true, count, datatype, rpeer, schedule, true);
+      res = NBC_Sched_recv (rbuf, tmprbuf, count, datatype, rpeer, schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;
       }
 
-      res = NBC_Sched_op (recvbuf, false, 0, true, recvbuf, false, count, datatype, op,
-                          schedule, true);
+      res = NBC_Sched_op (lbuf, tmplbuf, rbuf, tmprbuf, count, datatype, op, schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;
       }
+      /* swap left and right buffers */
+      buf = rbuf; rbuf = lbuf ; lbuf = buf;
+      tmprbuf ^= 1; tmplbuf ^= 1;
     }
 
     /* exchange our result with the remote root (each root will broadcast to the other's peers) */
-    res = NBC_Sched_recv (0, true, count, datatype, 0, schedule, false);
+    res = NBC_Sched_recv ((void *)(-gap), true, count, datatype, 0, schedule, false);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
       return res;
     }
@@ -643,7 +692,7 @@ static inline int allred_sched_linear(int rank, int rsize, const void *sendbuf,
 
     /* broadcast the result to all remote peers */
     for (int rpeer = 1 ; rpeer < rsize ; ++rpeer) {
-      res = NBC_Sched_send (0, true, count, datatype, rpeer, schedule, false);
+      res = NBC_Sched_send ((void *)(-gap), true, count, datatype, rpeer, schedule, false);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;
       }
diff --git a/ompi/mca/coll/libnbc/nbc_iexscan.c b/ompi/mca/coll/libnbc/nbc_iexscan.c
index 736f16c50b..73946e7dfa 100644
--- a/ompi/mca/coll/libnbc/nbc_iexscan.c
+++ b/ompi/mca/coll/libnbc/nbc_iexscan.c
@@ -45,7 +45,7 @@ int ompi_coll_libnbc_iexscan(const void* sendbuf, void* recvbuf, int count, MPI_
                              struct ompi_communicator_t *comm, ompi_request_t ** request,
                              struct mca_coll_base_module_2_1_0_t *module) {
     int rank, p, res;
-    MPI_Aint ext;
+    ptrdiff_t gap, span;
     NBC_Schedule *schedule;
 #ifdef NBC_CACHE_SCHEDULE
     NBC_Scan_args *args, *found, search;
@@ -59,27 +59,27 @@ int ompi_coll_libnbc_iexscan(const void* sendbuf, void* recvbuf, int count, MPI_
     rank = ompi_comm_rank (comm);
     p = ompi_comm_size (comm);
 
-    res = ompi_datatype_type_extent(datatype, &ext);
-    if (MPI_SUCCESS != res) {
-        NBC_Error("MPI Error in ompi_datatype_type_extent() (%i)", res);
-        return res;
-    }
-
     res = NBC_Init_handle(comm, &handle, libnbc_module);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;
     }
 
-    if (inplace && rank < p - 1) {
-        /* need more buffer space for the inplace case */
-        handle->tmpbuf = malloc(ext * count * 2);
-    } else {
-        handle->tmpbuf = malloc(ext * count);
-    }
-
-    if (handle->tmpbuf == NULL) {
-        NBC_Return_handle (handle);
-        return OMPI_ERR_OUT_OF_RESOURCE;
+    span = opal_datatype_span(&datatype->super, count, &gap);
+    if (0 < rank) {
+        handle->tmpbuf = malloc(span);
+        if (handle->tmpbuf == NULL) {
+            NBC_Return_handle (handle);
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        if (inplace) {
+            NBC_Copy(recvbuf, count, datatype, (char *)handle->tmpbuf-gap, count, datatype, comm);
+        } else {
+            NBC_Copy(sendbuf, count, datatype, (char *)handle->tmpbuf-gap, count, datatype, comm);
+        }
+        if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+            NBC_Return_handle (handle);
+            return res;
+        }
     }
 
 #ifdef NBC_CACHE_SCHEDULE
@@ -102,13 +102,7 @@ int ompi_coll_libnbc_iexscan(const void* sendbuf, void* recvbuf, int count, MPI_
         handle->schedule = schedule;
 
         if (rank != 0) {
-            if (inplace && rank < p - 1) {
-                /* if sendbuf == recvbuf do not clobber the send buffer until it has been combined
-                 * with the incoming data. */
-                res = NBC_Sched_recv ((void *) (ext * count), true, count, datatype, rank-1, schedule, false);
-            } else {
-                res = NBC_Sched_recv (recvbuf, false, count, datatype, rank-1, schedule, false);
-            }
+            res = NBC_Sched_recv (recvbuf, false, count, datatype, rank-1, schedule, false);
 
             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
                 NBC_Return_handle (handle);
@@ -123,15 +117,8 @@ int ompi_coll_libnbc_iexscan(const void* sendbuf, void* recvbuf, int count, MPI_
                     return res;
                 }
 
-                /* perform the reduce in my temporary buffer */
-                /* this cannot be done until handle->tmpbuf is unused :-( so barrier after */
-                if (inplace) {
-                    res = NBC_Sched_op (0, true, sendbuf, false, (void *)(ext * count), true, count,
-                                        datatype, op, schedule, true);
-                } else {
-                    res = NBC_Sched_op (0, true, sendbuf, false, recvbuf, false, count, datatype, op,
-                                        schedule, true);
-                }
+                res = NBC_Sched_op (recvbuf, false, (void *)(-gap), true, count,
+                                     datatype, op, schedule, true);
 
                 if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
                     NBC_Return_handle (handle);
@@ -139,24 +126,18 @@ int ompi_coll_libnbc_iexscan(const void* sendbuf, void* recvbuf, int count, MPI_
                 }
 
                 /* send reduced data onward */
-                res = NBC_Sched_send (0, true, count, datatype, rank + 1, schedule, false);
+                res = NBC_Sched_send ((void *)(-gap), true, count, datatype, rank + 1, schedule, false);
                 if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
                     NBC_Return_handle (handle);
                     return res;
                 }
-
-                if (inplace) {
-                    /* copy the received data into the receive buffer */
-                    res = NBC_Sched_copy ((void *)(ext * count), true, count, datatype, recvbuf,
-                                          false, count, datatype, schedule, false);
-                    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-                        NBC_Return_handle (handle);
-                        return res;
-                    }
-                }
             }
         } else if (p > 1) {
-            res = NBC_Sched_send (sendbuf, false, count, datatype, 1, schedule, false);
+            if (inplace) {
+              res = NBC_Sched_send (recvbuf, false, count, datatype, 1, schedule, false);
+            } else {
+              res = NBC_Sched_send (sendbuf, false, count, datatype, 1, schedule, false);
+            }
             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
                 NBC_Return_handle (handle);
                 return res;
diff --git a/ompi/mca/coll/libnbc/nbc_internal.h b/ompi/mca/coll/libnbc/nbc_internal.h
index 7b7e3210f9..76a248d3d9 100644
--- a/ompi/mca/coll/libnbc/nbc_internal.h
+++ b/ompi/mca/coll/libnbc/nbc_internal.h
@@ -10,7 +10,7 @@
  *
  * Copyright (c) 2012      Oracle and/or its affiliates.  All rights reserved.
  * Copyright (c) 2014      NVIDIA Corporation.  All rights reserved.
- * Copyright (c) 2015      Research Organization for Information Science
+ * Copyright (c) 2015-2016 Research Organization for Information Science
  *                         and Technology (RIST). All rights reserved.
  * Copyright (c) 2015      Los Alamos National Security, LLC. All rights
  *                         reserved.
@@ -92,6 +92,7 @@ typedef struct {
   MPI_Datatype datatype;
   int dest;
   char tmpbuf;
+  bool local;
 } NBC_Args_send;
 
 /* the receive argument struct */
@@ -102,6 +103,7 @@ typedef struct {
   MPI_Datatype datatype;
   char tmpbuf;
   int source;
+  bool local;
 } NBC_Args_recv;
 
 /* the operation argument struct */
@@ -109,10 +111,8 @@ typedef struct {
   NBC_Fn_type type;
   char tmpbuf1;
   char tmpbuf2;
-  char tmpbuf3;
   const void *buf1;
   void *buf2;
-  void *buf3;
   MPI_Op op;
   MPI_Datatype datatype;
   int count;
@@ -144,8 +144,10 @@ typedef struct {
 
 /* internal function prototypes */
 int NBC_Sched_send (const void* buf, char tmpbuf, int count, MPI_Datatype datatype, int dest, NBC_Schedule *schedule, bool barrier);
+int NBC_Sched_local_send (const void* buf, char tmpbuf, int count, MPI_Datatype datatype, int dest,NBC_Schedule *schedule, bool barrier);
 int NBC_Sched_recv (void* buf, char tmpbuf, int count, MPI_Datatype datatype, int source, NBC_Schedule *schedule, bool barrier);
-int NBC_Sched_op (void* buf3, char tmpbuf3, const void* buf1, char tmpbuf1, void* buf2, char tmpbuf2, int count, MPI_Datatype datatype,
+int NBC_Sched_local_recv (void* buf, char tmpbuf, int count, MPI_Datatype datatype, int source, NBC_Schedule *schedule, bool barrier);
+int NBC_Sched_op (const void* buf1, char tmpbuf1, void* buf2, char tmpbuf2, int count, MPI_Datatype datatype,
                   MPI_Op op, NBC_Schedule *schedule, bool barrier);
 int NBC_Sched_copy (void *src, char tmpsrc, int srccount, MPI_Datatype srctype, void *tgt, char tmptgt, int tgtcount,
                     MPI_Datatype tgttype, NBC_Schedule *schedule, bool barrier);
diff --git a/ompi/mca/coll/libnbc/nbc_ireduce.c b/ompi/mca/coll/libnbc/nbc_ireduce.c
index 0045deb6a5..387988a4f5 100644
--- a/ompi/mca/coll/libnbc/nbc_ireduce.c
+++ b/ompi/mca/coll/libnbc/nbc_ireduce.c
@@ -7,20 +7,23 @@
  *                         rights reserved.
  * Copyright (c) 2013-2015 Los Alamos National Security, LLC. All rights
  *                         reserved.
- * Copyright (c) 2014-2015 Research Organization for Information Science
+ * Copyright (c) 2014-2016 Research Organization for Information Science
  *                         and Technology (RIST). All rights reserved.
  *
  * Author(s): Torsten Hoefler <htor@cs.indiana.edu>
  *
  */
+
+#include "ompi/op/op.h"
+
 #include "nbc_internal.h"
 
-static inline int red_sched_binomial (int rank, int p, int root, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
-                                      MPI_Op op, void *redbuf, NBC_Schedule *schedule, NBC_Handle *handle);
+static inline int red_sched_binomial (int rank, int p, int root, const void *sendbuf, void *redbuf, int count, MPI_Datatype datatype,
+                                      MPI_Op op, char inplace, NBC_Schedule *schedule, NBC_Handle *handle);
 static inline int red_sched_chain (int rank, int p, int root, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
                                    MPI_Op op, int ext, size_t size, NBC_Schedule *schedule, NBC_Handle *handle, int fragsize);
 
-static inline int red_sched_linear (int rank, int rsize, int root, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
+static inline int red_sched_linear (int rank, int rsize, int root, const void *sendbuf, void *recvbuf, void *tmpbuf, int count, MPI_Datatype datatype,
                                     MPI_Op op, NBC_Schedule *schedule, NBC_Handle *handle);
 
 #ifdef NBC_CACHE_SCHEDULE
@@ -55,6 +58,7 @@ int ompi_coll_libnbc_ireduce(const void* sendbuf, void* recvbuf, int count, MPI_
   enum { NBC_RED_BINOMIAL, NBC_RED_CHAIN } alg;
   NBC_Handle *handle;
   ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
+  ptrdiff_t span, gap;
 
   NBC_IN_PLACE(sendbuf, recvbuf, inplace);
 
@@ -89,19 +93,22 @@ int ompi_coll_libnbc_ireduce(const void* sendbuf, void* recvbuf, int count, MPI_
     return res;
   }
 
+  span = opal_datatype_span(&datatype->super, count, &gap);
+
   /* algorithm selection */
-  if (p > 4 || size * count < 65536) {
+  if (p > 4 || size * count < 65536 || !ompi_op_is_commute(op)) {
     alg = NBC_RED_BINOMIAL;
     if(rank == root) {
       /* root reduces in receivebuffer */
-      handle->tmpbuf = malloc (ext * count);
+      handle->tmpbuf = malloc (span);
+      redbuf = recvbuf;
     } else {
       /* recvbuf may not be valid on non-root nodes */
-      handle->tmpbuf = malloc (ext * count * 2);
-      redbuf = (char*) handle->tmpbuf + ext * count;
+      handle->tmpbuf = malloc (2*span);
+      redbuf = (char*) handle->tmpbuf + span - gap;
     }
   } else {
-    handle->tmpbuf = malloc (ext * count);
+    handle->tmpbuf = malloc (span);
     alg = NBC_RED_CHAIN;
     segsize = 16384/2;
   }
@@ -135,7 +142,7 @@ int ompi_coll_libnbc_ireduce(const void* sendbuf, void* recvbuf, int count, MPI_
 
     switch(alg) {
       case NBC_RED_BINOMIAL:
-        res = red_sched_binomial(rank, p, root, sendbuf, recvbuf, count, datatype, op, redbuf, schedule, handle);
+        res = red_sched_binomial(rank, p, root, sendbuf, redbuf, count, datatype, op, inplace, schedule, handle);
         break;
       case NBC_RED_CHAIN:
         res = red_sched_chain(rank, p, root, sendbuf, recvbuf, count, datatype, op, ext, size, schedule, handle, segsize);
@@ -201,25 +208,20 @@ int ompi_coll_libnbc_ireduce_inter(const void* sendbuf, void* recvbuf, int count
 				   struct mca_coll_base_module_2_1_0_t *module) {
   int rank, res, rsize;
   NBC_Schedule *schedule;
-  MPI_Aint ext;
   NBC_Handle *handle;
   ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
+  ptrdiff_t span, gap;
 
   rank = ompi_comm_rank (comm);
   rsize = ompi_comm_remote_size (comm);
 
-  res = ompi_datatype_type_extent (datatype, &ext);
-  if (MPI_SUCCESS != res) {
-    NBC_Error("MPI Error in ompi_datatype_type_extent() (%i)", res);
-    return res;
-  }
-
   res = NBC_Init_handle(comm, &handle, libnbc_module);
   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
     return res;
   }
 
-  handle->tmpbuf = malloc (ext * count);
+  span = opal_datatype_span(&datatype->super, count, &gap);
+  handle->tmpbuf = malloc (span);
   if (OPAL_UNLIKELY(NULL == handle->tmpbuf)) {
     NBC_Return_handle (handle);
     return OMPI_ERR_OUT_OF_RESOURCE;
@@ -231,7 +233,7 @@ int ompi_coll_libnbc_ireduce_inter(const void* sendbuf, void* recvbuf, int count
     return OMPI_ERR_OUT_OF_RESOURCE;
   }
 
-  res = red_sched_linear (rank, rsize, root, sendbuf, recvbuf, count, datatype, op, schedule, handle);
+  res = red_sched_linear (rank, rsize, root, sendbuf, recvbuf, (void *)(-gap), count, datatype, op, schedule, handle);
   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
     NBC_Return_handle (handle);
     return OMPI_ERR_OUT_OF_RESOURCE;
@@ -257,6 +259,8 @@ int ompi_coll_libnbc_ireduce_inter(const void* sendbuf, void* recvbuf, int count
 
 
 /* binomial reduce
+ * if op is not commutative, reduce on rank 0, and then send the result to root rank
+ *
  * working principle:
  * - each node gets a virtual rank vrank
  * - the 'root' node get vrank 0
@@ -285,65 +289,81 @@ int ompi_coll_libnbc_ireduce_inter(const void* sendbuf, void* recvbuf, int count
   if (vrank == 0) rank = root; \
   if (vrank == root) rank = 0; \
 }
-static inline int red_sched_binomial (int rank, int p, int root, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
-                                      MPI_Op op, void *redbuf, NBC_Schedule *schedule, NBC_Handle *handle) {
-  int vrank, vpeer, peer, res, maxr;
+static inline int red_sched_binomial (int rank, int p, int root, const void *sendbuf, void *redbuf, int count, MPI_Datatype datatype,
+                                      MPI_Op op, char inplace, NBC_Schedule *schedule, NBC_Handle *handle) {
+  int vroot, vrank, vpeer, peer, res, maxr;
+  char *rbuf, *lbuf, *buf;
+  int tmprbuf, tmplbuf;
+  ptrdiff_t gap;
+  (void)opal_datatype_span(&datatype->super, count, &gap);
 
-  RANK2VRANK(rank, vrank, root);
+  if (ompi_op_is_commute(op)) {
+    vroot = root;
+  } else {
+    vroot = 0;
+  }
+  RANK2VRANK(rank, vrank, vroot);
   maxr = (int)ceil((log((double)p)/LOG2));
 
+  /* ensure the result ends up in redbuf on vrank 0 */
+  if (0 == (maxr%2)) {
+    rbuf = (void *)(-gap);
+    tmprbuf = true;
+    lbuf = redbuf;
+    tmplbuf = false;
+  } else {
+    lbuf = (void *)(-gap);
+    tmplbuf = true;
+    rbuf = redbuf;
+    tmprbuf = false;
+    if (inplace) {
+        res = NBC_Copy(rbuf, count, datatype, ((char *)handle->tmpbuf)-gap, count, datatype, MPI_COMM_SELF);
+        if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+          return res;
+        }
+    }
+  }
+
   for (int r = 1, firstred = 1 ; r <= maxr ; ++r) {
     if ((vrank % (1 << r)) == 0) {
       /* we have to receive this round */
       vpeer = vrank + (1 << (r - 1));
-      VRANK2RANK(peer, vpeer, root)
+      VRANK2RANK(peer, vpeer, vroot)
       if (peer < p) {
         /* we have to wait until we have the data */
-        res = NBC_Sched_recv (0, true, count, datatype, peer, schedule, true);
+        res = NBC_Sched_recv (rbuf, tmprbuf, count, datatype, peer, schedule, true);
         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
           return res;
         }
 
         /* perform the reduce in my local buffer */
         /* this cannot be done until handle->tmpbuf is unused :-( so barrier after the op */
-        if (firstred) {
-          if (rank == root) {
-            /* root is the only one who reduces in the receivebuffer
-             * take data from sendbuf in first round - save copy */
-            res = NBC_Sched_op (recvbuf, false, sendbuf, false, 0, true, count, datatype, op, schedule, true);
-          } else {
-            /* all others may not have a receive buffer
-             * take data from sendbuf in first round - save copy */
-            res = NBC_Sched_op ((char *) redbuf - (intptr_t) handle->tmpbuf, true, sendbuf, false, 0, true, count,
-                                datatype, op, schedule, true);
-          }
+        if (firstred && !inplace) {
+          /* perform the reduce with the senbuf */
+          res = NBC_Sched_op (sendbuf, false, rbuf, tmprbuf, count, datatype, op, schedule, true);
           firstred = 0;
         } else {
-          if(rank == root) {
-            /* root is the only one who reduces in the receivebuffer */
-            res = NBC_Sched_op (recvbuf, false, recvbuf, false, 0, true, count, datatype, op, schedule, true);
-          } else {
-            /* all others may not have a receive buffer */
-            res = NBC_Sched_op ((char *) redbuf - (intptr_t) handle->tmpbuf, true, (char *) redbuf - (intptr_t) handle->tmpbuf,
-                                true, 0, true, count, datatype, op, schedule, true);
-          }
+          /* perform the reduce in my local buffer */
+          res = NBC_Sched_op (lbuf, tmplbuf, rbuf, tmprbuf, count, datatype, op, schedule, true);
         }
 
         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
           return res;
         }
+        /* swap left and right buffers */
+        buf = rbuf; rbuf = lbuf ; lbuf = buf;
+        tmprbuf ^= 1; tmplbuf ^= 1;
       }
     } else {
       /* we have to send this round */
       vpeer = vrank - (1 << (r - 1));
-      VRANK2RANK(peer, vpeer, root)
-      if (firstred) {
-        /* we did not reduce anything */
+      VRANK2RANK(peer, vpeer, vroot)
+      if (firstred && !inplace) {
+        /* we have to use the sendbuf in the first round .. */
         res = NBC_Sched_send (sendbuf, false, count, datatype, peer, schedule, false);
       } else {
-        /* we have to use the redbuf the root (which works in receivebuf) is never sending .. */
-        res = NBC_Sched_send ((char *) redbuf - (intptr_t) handle->tmpbuf, true, count, datatype, peer, schedule,
-                              false);
+        /* and the redbuf in all remaining rounds */
+        res = NBC_Sched_send (lbuf, tmplbuf, count, datatype, peer, schedule, false);
       }
 
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -354,6 +374,14 @@ static inline int red_sched_binomial (int rank, int p, int root, const void *sen
       break;
     }
   }
+  /* send to root if vroot ! root */
+  if (vroot != root) {
+    if (0 == rank) {
+      res = NBC_Sched_send (redbuf, false, count, datatype, root, schedule, false);
+    } else if (root == rank) {
+      res = NBC_Sched_recv (redbuf, false, count, datatype, vroot, schedule, false);
+    }
+  }
 
   return OMPI_SUCCESS;
 }
@@ -389,18 +417,22 @@ static inline int red_sched_chain (int rank, int p, int root, const void *sendbu
 
     /* last node does not recv */
     if (vrank != p-1) {
-      res = NBC_Sched_recv ((char *) offset, true, thiscount, datatype, rpeer, schedule, true);
+      if (vrank == 0) {
+        res = NBC_Sched_recv ((char *)recvbuf+offset, false, thiscount, datatype, rpeer, schedule, true);
+      } else {
+        res = NBC_Sched_recv ((char *) offset, true, thiscount, datatype, rpeer, schedule, true);
+      }
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;
       }
 
       /* root reduces into receivebuf */
       if(vrank == 0) {
-        res = NBC_Sched_op ((char *) recvbuf + offset, false, (char *) sendbuf + offset, false, (char *) offset, true,
-                            thiscount, datatype, op, schedule, true);
+        res = NBC_Sched_op ((char *) sendbuf + offset, false, (char *) recvbuf + offset, false,
+                             thiscount, datatype, op, schedule, true);
       } else {
-        res = NBC_Sched_op ((char *) offset, true, (char *) sendbuf + offset, false, (char *) offset, true, thiscount,
-                            datatype, op, schedule, true);
+        res = NBC_Sched_op ((char *) sendbuf + offset, false, (char *) offset, true, thiscount,
+                             datatype, op, schedule, true);
       }
 
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -428,33 +460,51 @@ static inline int red_sched_chain (int rank, int p, int root, const void *sendbu
 }
 
 /* simple linear algorithm for intercommunicators */
-static inline int red_sched_linear (int rank, int rsize, int root, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
+static inline int red_sched_linear (int rank, int rsize, int root, const void *sendbuf, void *recvbuf, void *tmpbuf, int count, MPI_Datatype datatype,
                                     MPI_Op op, NBC_Schedule *schedule, NBC_Handle *handle) {
   int res;
+  char *rbuf, *lbuf, *buf;
+  int tmprbuf, tmplbuf;
 
   if (0 == count) {
     return OMPI_SUCCESS;
   }
 
   if (MPI_ROOT == root) {
-    res = NBC_Sched_recv (recvbuf, false, count, datatype, 0, schedule, true);
+    /* ensure the result ends up in recvbuf */
+    if (0 == (rsize%2)) {
+      lbuf = tmpbuf;
+      tmplbuf = true;
+      rbuf = recvbuf;
+      tmprbuf = false;
+    } else {
+      rbuf = tmpbuf;
+      tmprbuf = true;
+      lbuf = recvbuf;
+      tmplbuf = false;
+    }
+
+    res = NBC_Sched_recv (lbuf, tmplbuf, count, datatype, 0, schedule, false);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
       return res;
     }
 
     for (int peer = 1 ; peer < rsize ; ++peer) {
-      res = NBC_Sched_recv (0, true, count, datatype, peer, schedule, true);
+      res = NBC_Sched_recv (rbuf, tmprbuf, count, datatype, peer, schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;
       }
 
-      res = NBC_Sched_op (recvbuf, false, 0, true, recvbuf, false, count, datatype, op, schedule, true);
+      res = NBC_Sched_op (lbuf, tmplbuf, rbuf, tmprbuf, count, datatype, op, schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;
       }
+      /* swap left and right buffers */
+      buf = rbuf; rbuf = lbuf ; lbuf = buf;
+      tmprbuf ^= 1; tmplbuf ^= 1;
     }
   } else if (MPI_PROC_NULL != root) {
-    res = NBC_Sched_send (sendbuf, false, count, datatype, root, schedule, false);
+    res = NBC_Sched_send (sendbuf, false, count, datatype, root, 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 cd1dad14e7..c5fb0fdbbb 100644
--- a/ompi/mca/coll/libnbc/nbc_ireduce_scatter.c
+++ b/ompi/mca/coll/libnbc/nbc_ireduce_scatter.c
@@ -7,7 +7,7 @@
  *                         rights reserved.
  * Copyright (c) 2013-2015 Los Alamos National Security, LLC. All rights
  *                         reserved.
- * Copyright (c) 2014-2015 Research Organization for Information Science
+ * Copyright (c) 2014-2016 Research Organization for Information Science
  *                         and Technology (RIST). All rights reserved.
  * Copyright (c) 2015      The University of Tennessee and The University
  *                         of Tennessee Research Foundation. All rights
@@ -40,10 +40,12 @@ int ompi_coll_libnbc_ireduce_scatter(const void* sendbuf, void* recvbuf, const i
                                      struct mca_coll_base_module_2_1_0_t *module) {
   int peer, rank, maxr, p, res, count;
   MPI_Aint ext;
-  char *redbuf, *sbuf, inplace;
+  ptrdiff_t gap, span;
+  char *sbuf, inplace;
   NBC_Schedule *schedule;
   NBC_Handle *handle;
   ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
+  char *rbuf, *lbuf, *buf;
 
   NBC_IN_PLACE(sendbuf, recvbuf, inplace);
 
@@ -81,13 +83,15 @@ int ompi_coll_libnbc_ireduce_scatter(const void* sendbuf, void* recvbuf, const i
 
   maxr = (int) ceil ((log((double) p) / LOG2));
 
-  handle->tmpbuf = malloc (ext * count * 2);
+  span = opal_datatype_span(&datatype->super, count, &gap);
+  handle->tmpbuf = malloc (span * 2);
   if (OPAL_UNLIKELY(NULL == handle->tmpbuf)) {
     NBC_Return_handle (handle);
     return OMPI_ERR_OUT_OF_RESOURCE;
   }
 
-  redbuf = (char *) handle->tmpbuf + ext * count;
+  rbuf = (char *)(-gap);
+  lbuf = (char *)(span - gap);
 
   schedule = OBJ_NEW(NBC_Schedule);
   if (OPAL_UNLIKELY(NULL == schedule)) {
@@ -104,7 +108,7 @@ int ompi_coll_libnbc_ireduce_scatter(const void* sendbuf, void* recvbuf, const i
       peer = rank + (1 << (r - 1));
       if (peer < p) {
         /* we have to wait until we have the data */
-        res = NBC_Sched_recv(0, true, count, datatype, peer, schedule, true);
+        res = NBC_Sched_recv(rbuf, true, count, datatype, peer, schedule, true);
         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
           NBC_Return_handle (handle);
           return res;
@@ -113,19 +117,19 @@ int ompi_coll_libnbc_ireduce_scatter(const void* sendbuf, void* recvbuf, const i
         /* this cannot be done until handle->tmpbuf is unused :-( so barrier after the op */
         if (firstred) {
           /* take reduce data from the sendbuf in the first round -> save copy */
-          res = NBC_Sched_op (redbuf - (intptr_t) handle->tmpbuf, true, sendbuf, false, 0, true, count, datatype,
-                              op, schedule, true);
+          res = NBC_Sched_op (sendbuf, false, rbuf, true, count, datatype, op, schedule, true);
           firstred = 0;
         } else {
           /* perform the reduce in my local buffer */
-          res = NBC_Sched_op (redbuf - (intptr_t) handle->tmpbuf, true, redbuf - (intptr_t) handle->tmpbuf, true,
-                              0, true, count, datatype, op, schedule, true);
+          res = NBC_Sched_op (lbuf, true, rbuf, true, count, datatype, op, schedule, true);
         }
 
         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
           NBC_Return_handle (handle);
           return res;
         }
+        /* swap left and right buffers */
+        buf = rbuf; rbuf = lbuf ; lbuf = buf;
       }
     } else {
       /* we have to send this round */
@@ -134,8 +138,8 @@ int ompi_coll_libnbc_ireduce_scatter(const void* sendbuf, void* recvbuf, const i
         /* we have to send the senbuf */
         res = NBC_Sched_send (sendbuf, false, count, datatype, peer, schedule, false);
       } else {
-        /* we send an already reduced value from redbuf */
-        res = NBC_Sched_send (redbuf - (intptr_t) handle->tmpbuf, true, count, datatype, peer, schedule, false);
+        /* we send an already reduced value from lbuf */
+        res = NBC_Sched_send (lbuf, true, count, datatype, peer, schedule, false);
       }
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         NBC_Return_handle (handle);
@@ -157,9 +161,9 @@ int ompi_coll_libnbc_ireduce_scatter(const void* sendbuf, void* recvbuf, const i
   if (rank == 0) {
     for (long int r = 1, offset = 0 ; r < p ; ++r) {
       offset += recvcounts[r-1];
-      sbuf = redbuf + offset * ext;
+      sbuf = lbuf + (offset*ext);
       /* root sends the right buffer to the right receiver */
-      res = NBC_Sched_send (sbuf - (intptr_t) handle->tmpbuf, true, recvcounts[r], datatype, r, schedule,
+      res = NBC_Sched_send (sbuf, true, recvcounts[r], datatype, r, schedule,
                             false);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         NBC_Return_handle (handle);
@@ -167,7 +171,7 @@ int ompi_coll_libnbc_ireduce_scatter(const void* sendbuf, void* recvbuf, const i
       }
     }
 
-    res = NBC_Sched_copy (redbuf - (intptr_t) handle->tmpbuf, true, recvcounts[0], datatype, recvbuf, false,
+    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);
@@ -199,13 +203,15 @@ int ompi_coll_libnbc_ireduce_scatter(const void* sendbuf, void* recvbuf, const i
 int ompi_coll_libnbc_ireduce_scatter_inter (const void* sendbuf, void* recvbuf, const int *recvcounts, MPI_Datatype datatype,
                                             MPI_Op op, struct ompi_communicator_t *comm, ompi_request_t ** request,
                                             struct mca_coll_base_module_2_1_0_t *module) {
-  int rank, res, count, rsize;
+  int rank, res, count, lsize, rsize;
   MPI_Aint ext;
+  ptrdiff_t gap, span;
   NBC_Schedule *schedule;
   NBC_Handle *handle;
   ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
 
   rank = ompi_comm_rank (comm);
+  lsize = ompi_comm_size(comm);
   rsize = ompi_comm_remote_size (comm);
 
   res = ompi_datatype_type_extent (datatype, &ext);
@@ -215,17 +221,19 @@ int ompi_coll_libnbc_ireduce_scatter_inter (const void* sendbuf, void* recvbuf,
   }
 
   count = 0;
-  for (int r = 0 ; r < rsize ; ++r) {
+  for (int r = 0 ; r < lsize ; ++r) {
     count += recvcounts[r];
   }
 
+  span = opal_datatype_span(&datatype->super, count, &gap);
+
   res = NBC_Init_handle(comm, &handle, libnbc_module);
   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
     return res;
   }
 
   if (count > 0) {
-    handle->tmpbuf = malloc (2 * ext * count);
+    handle->tmpbuf = malloc (2 * span);
     if (OPAL_UNLIKELY(NULL == handle->tmpbuf)) {
       NBC_Return_handle (handle);
       return OMPI_ERR_OUT_OF_RESOURCE;
@@ -249,44 +257,55 @@ int ompi_coll_libnbc_ireduce_scatter_inter (const void* sendbuf, void* recvbuf,
   }
 
   if (0 == rank) {
-    res = NBC_Sched_recv ((void *) 0, true, count, datatype, 0, schedule, true);
+    char *lbuf, *rbuf;
+    lbuf = (char *)(-gap);
+    rbuf = (char *)(span-gap);
+    res = NBC_Sched_recv (lbuf, true, count, datatype, 0, schedule, true);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
       NBC_Return_handle (handle);
       return res;
     }
 
     for (int peer = 1 ; peer < rsize ; ++peer) {
-      res = NBC_Sched_recv ((void *)(ext * count), true, count, datatype, peer, schedule, true);
+      char *tbuf;
+      res = NBC_Sched_recv (rbuf, true, count, datatype, peer, schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         NBC_Return_handle (handle);
         return res;
       }
 
-      res = NBC_Sched_op ((void *) 0, true, (void *)(ext * count), true, (void *) 0, true, count, datatype,
+      res = NBC_Sched_op (lbuf, true, rbuf, true, count, datatype,
                           op, schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         NBC_Return_handle (handle);
         return res;
       }
+      tbuf = lbuf; lbuf = rbuf; rbuf = tbuf;
     }
 
     /* exchange data with remote root for scatter phase (we *could* use the local communicator to do the scatter) */
-    res = NBC_Sched_recv ((void *)(ext * count), true, count, datatype, 0, schedule, false);
+    res = NBC_Sched_recv (rbuf, true, count, datatype, 0, schedule, false);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
       NBC_Return_handle (handle);
       return res;
     }
 
-    res = NBC_Sched_send ((void *) 0, true, count, datatype, 0, schedule, true);
+    res = NBC_Sched_send (lbuf, true, count, datatype, 0, schedule, true);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
       NBC_Return_handle (handle);
       return res;
     }
 
-    /* scatter */
-    for (int peer = 0, offset = ext * count ; peer < rsize ; ++peer) {
-      res = NBC_Sched_send ((void *)(uintptr_t) offset, true, recvcounts[peer], datatype, peer, schedule,
-                            false);
+    /* do the local scatterv with the local communicator */
+    res = NBC_Sched_copy (lbuf, true, recvcounts[0], datatype, recvbuf, false,
+                          recvcounts[0], datatype, schedule, false);
+    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+      NBC_Return_handle (handle);
+      return res;
+    }
+    for (int peer = 1, offset = recvcounts[0] * ext; peer < lsize ; ++peer) {
+      res = NBC_Sched_local_send (lbuf + offset, true, recvcounts[peer], datatype, peer, schedule,
+                                  false);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         NBC_Return_handle (handle);
         return res;
@@ -294,13 +313,13 @@ int ompi_coll_libnbc_ireduce_scatter_inter (const void* sendbuf, void* recvbuf,
 
       offset += recvcounts[peer] * ext;
     }
-  }
-
-  /* receive my block */
-  res = NBC_Sched_recv (recvbuf, false, recvcounts[rank], datatype, 0, schedule, false);
-  if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-    NBC_Return_handle (handle);
-    return res;
+  } else {
+    /* receive my block */
+    res = NBC_Sched_local_recv (recvbuf, false, recvcounts[rank], datatype, 0, schedule, false);
+    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+      NBC_Return_handle (handle);
+      return res;
+    }
   }
 
   res = NBC_Sched_commit (schedule);
diff --git a/ompi/mca/coll/libnbc/nbc_ireduce_scatter_block.c b/ompi/mca/coll/libnbc/nbc_ireduce_scatter_block.c
index 4d553c6f55..f0c6721889 100644
--- a/ompi/mca/coll/libnbc/nbc_ireduce_scatter_block.c
+++ b/ompi/mca/coll/libnbc/nbc_ireduce_scatter_block.c
@@ -38,6 +38,7 @@ int ompi_coll_libnbc_ireduce_scatter_block(const void* sendbuf, void* recvbuf, i
                                            struct mca_coll_base_module_2_1_0_t *module) {
   int peer, rank, maxr, p, res, count;
   MPI_Aint ext;
+  ptrdiff_t gap, span;
   char *redbuf, *sbuf, inplace;
   NBC_Schedule *schedule;
   NBC_Handle *handle;
@@ -73,14 +74,19 @@ int ompi_coll_libnbc_ireduce_scatter_block(const void* sendbuf, void* recvbuf, i
   count = p * recvcount;
 
   if (0 < count) {
-    handle->tmpbuf = malloc (ext*count*2);
+    char *rbuf, *lbuf, *buf;
+
+    span = opal_datatype_span(&datatype->super, count, &gap);
+    handle->tmpbuf = malloc (2*span);
     if (NULL == handle->tmpbuf) {
       OMPI_COLL_LIBNBC_REQUEST_RETURN(handle);
       OBJ_RELEASE(schedule);
       return OMPI_ERR_OUT_OF_RESOURCE;
     }
 
-    redbuf = (char *) handle->tmpbuf + ext * count;
+    rbuf = (void *)(-gap);
+    lbuf = (char *)(span - gap);
+    redbuf = (char *) handle->tmpbuf + span - gap;
 
     /* copy data to redbuf if we only have a single node */
     if ((p == 1) && !inplace) {
@@ -98,7 +104,7 @@ int ompi_coll_libnbc_ireduce_scatter_block(const void* sendbuf, void* recvbuf, i
         peer = rank + (1 << (r - 1));
         if (peer < p) {
           /* we have to wait until we have the data */
-          res = NBC_Sched_recv (0, true, count, datatype, peer, schedule, true);
+          res = NBC_Sched_recv (rbuf, true, count, datatype, peer, schedule, true);
           if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
             NBC_Return_handle (handle);
             return res;
@@ -106,29 +112,29 @@ int ompi_coll_libnbc_ireduce_scatter_block(const void* sendbuf, void* recvbuf, i
 
           if (firstred) {
             /* take reduce data from the sendbuf in the first round -> save copy */
-            res = NBC_Sched_op (redbuf-(unsigned long)handle->tmpbuf, true, sendbuf, false, 0, true, count,
-                                datatype, op, schedule, true);
+            res = NBC_Sched_op (sendbuf, false, rbuf, true, count, datatype, op, schedule, true);
             firstred = 0;
           } else {
           /* perform the reduce in my local buffer */
-            res = NBC_Sched_op (redbuf-(unsigned long)handle->tmpbuf, true, redbuf-(unsigned long)handle->tmpbuf,
-                                true, 0, true, count, datatype, op, schedule, true);
+            res = NBC_Sched_op (lbuf, true, rbuf, true, count, datatype, op, schedule, true);
           }
 
           if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
             NBC_Return_handle (handle);
             return res;
           }
+          /* swap left and right buffers */
+          buf = rbuf; rbuf = lbuf ; lbuf = buf;
         }
       } else {
         /* we have to send this round */
         peer = rank - (1 << (r - 1));
         if(firstred) {
           /* we have to send the senbuf */
-          res = NBC_Sched_send (sendbuf, false, count, datatype, peer, schedule, true);
+          res = NBC_Sched_send (sendbuf, false, count, datatype, peer, schedule, false);
         } else {
           /* we send an already reduced value from redbuf */
-          res = NBC_Sched_send (redbuf-(unsigned long)handle->tmpbuf, true, count, datatype, peer, schedule, true);
+          res = NBC_Sched_send (lbuf, true, count, datatype, peer, schedule, false);
         }
 
         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -157,17 +163,19 @@ int ompi_coll_libnbc_ireduce_scatter_block(const void* sendbuf, void* recvbuf, i
     } else {
       for (int r = 1, offset = 0 ; r < p ; ++r) {
         offset += recvcount;
-        sbuf = ((char *)redbuf) + (offset*ext);
+        sbuf = lbuf + (offset*ext);
         /* root sends the right buffer to the right receiver */
-        res = NBC_Sched_send (sbuf-(unsigned long)handle->tmpbuf, true, recvcount, datatype, r, schedule, false);
+        res = NBC_Sched_send (sbuf, true, recvcount, datatype, r, schedule, false);
         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
           NBC_Return_handle (handle);
           return res;
         }
       }
 
-      res = NBC_Sched_copy (redbuf-(unsigned long)handle->tmpbuf, true, recvcount, datatype, recvbuf, false, recvcount,
-                            datatype, schedule, false);
+      if ((p != 1) || !inplace) {
+        res = NBC_Sched_copy (lbuf, true, recvcount, datatype, recvbuf, false, recvcount,
+                              datatype, schedule, false);
+      }
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         NBC_Return_handle (handle);
         return res;
@@ -193,16 +201,18 @@ int ompi_coll_libnbc_ireduce_scatter_block(const void* sendbuf, void* recvbuf, i
   return OMPI_SUCCESS;
 }
 
-int ompi_coll_libnbc_ireduce_scatter_block_inter(const void *sbuf, void *rbuf, int rcount, struct ompi_datatype_t *dtype,
+int ompi_coll_libnbc_ireduce_scatter_block_inter(const void *sendbuf, void *recvbuf, int rcount, struct ompi_datatype_t *dtype,
 						 struct ompi_op_t *op, struct ompi_communicator_t *comm,
 						 ompi_request_t **request, struct mca_coll_base_module_2_1_0_t *module) {
-  int rank, res, count, rsize;
+  int rank, res, count, lsize, rsize;
   MPI_Aint ext;
+  ptrdiff_t gap, span;
   NBC_Schedule *schedule;
   NBC_Handle *handle;
   ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
 
   rank = ompi_comm_rank (comm);
+  lsize = ompi_comm_size (comm);
   rsize = ompi_comm_remote_size (comm);
 
   res = ompi_datatype_type_extent (dtype, &ext);
@@ -216,10 +226,12 @@ int ompi_coll_libnbc_ireduce_scatter_block_inter(const void *sbuf, void *rbuf, i
     return res;
   }
 
-  count = rcount * rsize;
+  count = rcount * lsize;
+
+  span = opal_datatype_span(&dtype->super, count, &gap);
 
   if (count > 0) {
-    handle->tmpbuf = malloc (2 * ext * count);
+    handle->tmpbuf = malloc (2 * span);
     if (NULL == handle->tmpbuf) {
       NBC_Return_handle (handle);
       return OMPI_ERR_OUT_OF_RESOURCE;
@@ -236,62 +248,60 @@ int ompi_coll_libnbc_ireduce_scatter_block_inter(const void *sbuf, void *rbuf, i
   handle->schedule = schedule;
 
   /* send my data to the remote root */
-  res = NBC_Sched_send (sbuf, false, count, dtype, 0, schedule, false);
+  res = NBC_Sched_send (sendbuf, false, count, dtype, 0, schedule, false);
   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
     NBC_Return_handle (handle);
     return res;
   }
 
   if (0 == rank) {
-    res = NBC_Sched_recv ((void *) 0, true, count, dtype, 0, schedule, true);
+    char *lbuf, *rbuf;
+    lbuf = (char *)(-gap);
+    rbuf = (char *)(span-gap);
+    res = NBC_Sched_recv (lbuf, true, count, dtype, 0, schedule, true);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
       NBC_Return_handle (handle);
       return res;
     }
 
     for (int peer = 1 ; peer < rsize ; ++peer) {
-      res = NBC_Sched_recv ((void *)(ext * count), true, count, dtype, peer, schedule, true);
+      char *tbuf;
+      res = NBC_Sched_recv (rbuf, true, count, dtype, peer, schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         NBC_Return_handle (handle);
         return res;
       }
 
-      res = NBC_Sched_op ((void *) 0, true, (void *)(ext * count), true, (void *) 0, true, count, dtype, op,
-                          schedule, true);
+      res = NBC_Sched_op (lbuf, true, rbuf, true, count, dtype,
+                          op, schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         NBC_Return_handle (handle);
         return res;
       }
+      tbuf = lbuf; lbuf = rbuf; rbuf = tbuf;
     }
 
-    /* exchange data with remote root for scatter phase (we *could* use the local communicator to do the scatter) */
-    res = NBC_Sched_recv ((void *)(ext * count), true, count, dtype, 0, schedule, false);
+    /* do the scatter with the local communicator */
+    res = NBC_Sched_copy (lbuf, true, rcount, dtype, recvbuf, false, rcount,
+                          dtype, schedule, false);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
       NBC_Return_handle (handle);
       return res;
     }
-
-    res = NBC_Sched_send ((void *) 0, true, count, dtype, 0, schedule, true);
-    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-      NBC_Return_handle (handle);
-      return res;
-    }
-
-    /* scatter */
-    for (int peer = 0 ; peer < rsize ; ++peer) {
-      res = NBC_Sched_send ((void *)(ext * (count + peer * rcount)), true, rcount, dtype, peer, schedule, false);
+    for (int peer = 1 ; peer < lsize ; ++peer) {
+      res = NBC_Sched_local_send (lbuf + ext * rcount * peer, true, rcount, dtype, peer, schedule, false);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         NBC_Return_handle (handle);
         return res;
       }
     }
-  }
-
-  /* receive my block */
-  res = NBC_Sched_recv(rbuf, true, rcount, dtype, 0, schedule, false);
-  if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-    NBC_Return_handle (handle);
-    return res;
+  } else {
+    /* receive my block */
+    res = NBC_Sched_local_recv(recvbuf, false, rcount, dtype, 0, schedule, false);
+    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+      NBC_Return_handle (handle);
+      return res;
+    }
   }
 
   /*NBC_PRINT_SCHED(*schedule);*/
diff --git a/ompi/mca/coll/libnbc/nbc_iscan.c b/ompi/mca/coll/libnbc/nbc_iscan.c
index a239d14ed1..5b8b0bbdc1 100644
--- a/ompi/mca/coll/libnbc/nbc_iscan.c
+++ b/ompi/mca/coll/libnbc/nbc_iscan.c
@@ -36,16 +36,16 @@ int NBC_Scan_args_compare(NBC_Scan_args *a, NBC_Scan_args *b, void *param) {
 
 /* linear iscan
  * working principle:
- * 1. each node (but node 0) receives from left neigbor
+ * 1. each node (but node 0) receives from left neighbor
  * 2. performs op
- * 3. all but rank p-1 do sends to it's right neigbor and exits
+ * 3. all but rank p-1 do sends to it's right neighbor and exits
  *
  */
 int ompi_coll_libnbc_iscan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
                            struct ompi_communicator_t *comm, ompi_request_t ** request,
                            struct mca_coll_base_module_2_1_0_t *module) {
   int rank, p, res;
-  MPI_Aint ext;
+  ptrdiff_t gap, span;
   NBC_Schedule *schedule;
   char inplace;
   NBC_Handle *handle;
@@ -56,13 +56,7 @@ int ompi_coll_libnbc_iscan(const void* sendbuf, void* recvbuf, int count, MPI_Da
   rank = ompi_comm_rank (comm);
   p = ompi_comm_size (comm);
 
-  res = ompi_datatype_type_extent (datatype, &ext);
-  if (MPI_SUCCESS != res) {
-    NBC_Error("MPI Error in ompi_datatype_type_extent() (%i)", res);
-    return res;
-  }
-
-  if ((rank == 0) && !inplace) {
+  if (!inplace) {
     /* copy data to receivebuf */
     res = NBC_Copy (sendbuf, count, datatype, recvbuf, count, datatype, comm);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -75,12 +69,6 @@ int ompi_coll_libnbc_iscan(const void* sendbuf, void* recvbuf, int count, MPI_Da
     return res;
   }
 
-  handle->tmpbuf = malloc (ext * count);
-  if (NULL == handle->tmpbuf) {
-    NBC_Return_handle (handle);
-    return OMPI_ERR_OUT_OF_RESOURCE;
-  }
-
 #ifdef NBC_CACHE_SCHEDULE
   NBC_Scan_args *args, *found, search;
 
@@ -103,8 +91,15 @@ int ompi_coll_libnbc_iscan(const void* sendbuf, void* recvbuf, int count, MPI_Da
     handle->schedule = schedule;
 
     if(rank != 0) {
+      span = opal_datatype_span(&datatype->super, count, &gap);
+      handle->tmpbuf = malloc (span);
+      if (NULL == handle->tmpbuf) {
+        NBC_Return_handle (handle);
+        return OMPI_ERR_OUT_OF_RESOURCE;
+      }
+
       /* we have to wait until we have the data */
-      res = NBC_Sched_recv (0, true, count, datatype, rank-1, schedule, true);
+      res = NBC_Sched_recv ((void *)(-gap), true, count, datatype, rank-1, schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         NBC_Return_handle (handle);
         return res;
@@ -112,7 +107,7 @@ int ompi_coll_libnbc_iscan(const void* sendbuf, void* recvbuf, int count, MPI_Da
 
       /* perform the reduce in my local buffer */
       /* this cannot be done until handle->tmpbuf is unused :-( so barrier after the op */
-      res = NBC_Sched_op (recvbuf, false, sendbuf, false, 0, true, count, datatype, op, schedule,
+      res = NBC_Sched_op ((void *)(-gap), true, recvbuf, false, count, datatype, op, schedule,
                           true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         NBC_Return_handle (handle);