diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc
index 26f09c46c2224a4a46d302dff4b2ec594f0be103..a038bad701ba8ede3065af9f352f1f21784a50b7 100644
--- a/paddle/fluid/operators/listen_and_serv_op.cc
+++ b/paddle/fluid/operators/listen_and_serv_op.cc
@@ -27,6 +27,10 @@ limitations under the License. */
 #include "paddle/fluid/operators/distributed/request_handler_impl.h"
 #include "paddle/fluid/operators/listen_and_serv_op.h"
 
+DEFINE_int32(rpc_send_thread_num, 5, "number of threads for rpc send");
+DEFINE_int32(rpc_get_thread_num, 5, "number of threads for rpc get");
+DEFINE_int32(rpc_prefetch_thread_num, 5, "number of threads for rpc prefetch");
+
 namespace paddle {
 namespace operators {
 
@@ -332,11 +336,14 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
       sync_mode, checkpoint_block_id));
 
   rpc_service_->RegisterRPC(distributed::kRequestSend,
-                            request_send_handler_.get());
+                            request_send_handler_.get(),
+                            FLAGS_rpc_send_thread_num);
   rpc_service_->RegisterRPC(distributed::kRequestGet,
-                            request_get_handler_.get());
+                            request_get_handler_.get(),
+                            FLAGS_rpc_get_thread_num);
   rpc_service_->RegisterRPC(distributed::kRequestPrefetch,
-                            request_prefetch_handler_.get());
+                            request_prefetch_handler_.get(),
+                            FLAGS_rpc_prefetch_thread_num);
   rpc_service_->RegisterRPC(distributed::kRequestCheckpoint,
                             request_checkpoint_handler_.get());
 
diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py
index bcd4e4f6073eff1ea0449da8096030743158dd0f..737c8be8147a7efaf9b89827f063430146d3c078 100644
--- a/python/paddle/fluid/__init__.py
+++ b/python/paddle/fluid/__init__.py
@@ -121,6 +121,9 @@ def __bootstrap__():
         read_env_flags.append('rpc_server_profile_period')
         read_env_flags.append('rpc_server_profile_path')
         read_env_flags.append('enable_rpc_profiler')
+        read_env_flags.append('rpc_send_thread_num')
+        read_env_flags.append('rpc_get_thread_num')
+        read_env_flags.append('rpc_prefetch_thread_num')
 
     if core.is_compiled_with_cuda():
         read_env_flags += [