diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.h b/paddle/fluid/framework/details/multi_devices_graph_pass.h
index 26fc8dc198645c5cdb2eca8c2a02978d784e7b50..7cc68dd2d5a422cfa1ac3a4bfdd48545a6e5691d 100644
--- a/paddle/fluid/framework/details/multi_devices_graph_pass.h
+++ b/paddle/fluid/framework/details/multi_devices_graph_pass.h
@@ -132,8 +132,11 @@ class AsyncSSAGraphBuilder : public MultiDevSSAGraphBuilderBase {
       node->Op()->Flush();
     } else if (node->Name() == "lookup_table" || node->Name() == "nce" ||
                node->Name() == "hierarchical_sigmoid") {
+      // in async_mode, we do not need remote prefetch, because communicator
+      // will do async parameter recv.
       VLOG(1) << "set " << node->Name() << " op remote_prefetch to false";
       node->Op()->SetAttr("remote_prefetch", false);
+      node->Op()->Flush();
     }
     return false;
   }
diff --git a/paddle/fluid/framework/scope.h b/paddle/fluid/framework/scope.h
index cd752077d66f9b0292a634d42c64ecc4738c0007..6665458d4c8abd999385a7bf6bfd86c15e904b5f 100644
--- a/paddle/fluid/framework/scope.h
+++ b/paddle/fluid/framework/scope.h
@@ -52,6 +52,10 @@ class Scope {
   /// Mark it to const because that new kid scope cannot change parent scope.
   Scope& NewScope() const;
 
+  /// Create a sub-scope for current scope but do not record it in the kids to
+  /// avoid performance problems.
+  /// Note!!! You should delete the result pointer yourself to avoid memory
+  /// leak!
   Scope* NewTmpScope() const;
 
   /// Create a variable with given name if it doesn't exist.
diff --git a/paddle/fluid/operators/distributed/parameter_send.cc b/paddle/fluid/operators/distributed/parameter_send.cc
index ec2884c25290aa3cfd9818ead61119cc6c6b6feb..4858dbe84e05fe660c3b4d8fca19fffc49cee8f8 100644
--- a/paddle/fluid/operators/distributed/parameter_send.cc
+++ b/paddle/fluid/operators/distributed/parameter_send.cc
@@ -81,8 +81,8 @@ void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
     auto abs_sections = ToAbsoluteSection(rpc_ctx.height_sections);
 
     auto &send_rows = send_slr.rows();
-    std::vector<std::vector<int>> outs_rows_idx;
-    std::vector<std::vector<int>> outs_dense_idx;
+    std::vector<std::vector<size_t>> outs_rows_idx;
+    std::vector<std::vector<size_t>> outs_dense_idx;
 
     outs_rows_idx.resize(out_num);
     outs_dense_idx.resize(out_num);
@@ -99,7 +99,7 @@ void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
 
     // split rows index into output sparse vars
     for (size_t i = 0; i < send_rows.size(); ++i) {
-      int out_idx = GetSectionIndex(send_rows[i], abs_sections);
+      size_t out_idx = GetSectionIndex(send_rows[i], abs_sections);
       outs_rows_idx[out_idx].push_back(send_rows[i]);
       outs_dense_idx[out_idx].push_back(i);
     }
@@ -160,10 +160,9 @@ void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
     }
   }
 
-  // note!! only support sync send now
-  if (true || sync) {
-    for (size_t i = 0; i < rets.size(); i++) {
-      PADDLE_ENFORCE(rets[i]->Wait(), "internal error in RPCClient");
+  if (sync) {
+    for (auto &handle : rets) {
+      PADDLE_ENFORCE(handle->Wait(), "internal error in RPCClient");
     }
   }
 
diff --git a/paddle/fluid/operators/distributed_ops/send_op.cc b/paddle/fluid/operators/distributed_ops/send_op.cc
index 47688d0ad456873c93e9e7cdc1e550028347b052..b08cd0942f8c89b60d722c931d0cec2063b96578 100644
--- a/paddle/fluid/operators/distributed_ops/send_op.cc
+++ b/paddle/fluid/operators/distributed_ops/send_op.cc
@@ -52,7 +52,7 @@ class SendOp : public framework::OperatorBase {
         auto send_functor = distributed::ParameterSend<float>();
         auto rpc_ctx = distributed::RpcContext(ins[0], send_varnames, epmap,
                                                height_sections);
-        send_functor(rpc_ctx, scope, static_cast<bool>(sync_send));
+        send_functor(rpc_ctx, scope, true);
       } else {
         distributed::Communicator::GetInstance()->Send(ins[0], scope);
       }