DART-MPI: call into MPI every once in a while for local put/get

dart-impl/mpi/src/dart_communication.c

@@ -30,6 +30,11 @@
 #include <math.h>
 #include <alloca.h>
 
+/* the number of consecutive memcpy for local put/get before calling into MPI */
+#define NUM_CONSECUTIVE_MEMCPY 16
+
+/* number of performed local memcpy between calling into MPI */
+static _Thread_local int num_local_memcpy = 0;
 
 #define CHECK_UNITID_RANGE(_unitid, _team_data)                             \
   do {                                                                      \
@@ -220,13 +225,15 @@ dart__mpi__get_basic(
 {
   if (num_reqs) *num_reqs = 0;
 
-  if (team_data->unitid == team_unit_id.id) {
+  if (team_data->unitid == team_unit_id.id &&
+      num_local_memcpy++ < NUM_CONSECUTIVE_MEMCPY) {
     // use direct memcpy if we are on the same unit
     memcpy(dest, seginfo->selfbaseptr + offset,
         nelem * dart__mpi__datatype_sizeof(dtype));
     DART_LOG_DEBUG("dart_get: memcpy nelem:%zu "
         "source (coll.): offset:%lu -> dest: %p",
         nelem, offset, dest);
+    num_local_memcpy = 0;
     return DART_OK;
   }
 
@@ -354,12 +361,14 @@ dart__mpi__put_basic(
   if (num_reqs) *num_reqs = 0;
 
   /* copy data directly if we are on the same unit */
-  if (team_unit_id.id == team_data->unitid) {
+  if (team_unit_id.id == team_data->unitid &&
+      num_local_memcpy++ < NUM_CONSECUTIVE_MEMCPY) {
     if (flush_required_ptr) *flush_required_ptr = false;
     memcpy(seginfo->selfbaseptr + offset, src,
         nelem * dart__mpi__datatype_sizeof(dtype));
     DART_LOG_DEBUG("dart_put: memcpy nelem:%zu (from global allocation)"
         "offset: %"PRIu64"", nelem, offset);
+    num_local_memcpy = 0;
     return DART_OK;
   }