diff --git a/benchmark/cluster/vgg16/vgg16_fluid.py b/benchmark/cluster/vgg16/vgg16_fluid.py
index 05b5f3977cbed2f08df73c6d8ba2fff687db3313..0f5cd2a253554a7c37e1d34cd798f30ea2119bdd 100644
--- a/benchmark/cluster/vgg16/vgg16_fluid.py
+++ b/benchmark/cluster/vgg16/vgg16_fluid.py
@@ -38,7 +38,7 @@ def str2bool(v):
 
 parser = argparse.ArgumentParser(description=__doc__)
 parser.add_argument(
-    '--batch_size', type=int, default=128, help="Batch size for training.")
+    '--batch_size', type=int, default=16, help="Batch size for training.")
 parser.add_argument(
     '--learning_rate',
     type=float,
@@ -61,7 +61,7 @@ parser.add_argument(
 parser.add_argument(
     '--data_set',
     type=str,
-    default='cifar10',
+    default='flowers',
     choices=['cifar10', 'flowers'],
     help='Optional dataset for benchmark.')
 parser.add_argument(
@@ -200,26 +200,30 @@ def main():
                     fetch_list=[avg_cost, batch_acc, batch_size])
                 return loss, acc, b_size
 
-            if args.profile and args.task_index == 0:
-                # warmup.
-                for batch_id, data in enumerate(train_reader()):
-                    if batch_id > 5: break
-                    run_step(batch_id, data)
-                with profiler.profiler('All', 'total', '/tmp/profile_vgg'):
+            if args.profile:
+                with profiler.profiler('All', 'total',
+                                       '/tmp/profile_vgg_%d' % args.task_index):
                     for batch_id, data in enumerate(train_reader()):
-                        if batch_id > 5: break
+                        if batch_id > 4: break
                         run_step(batch_id, data)
 
+            total_time = 0.0
+            count = 0
             for batch_id, data in enumerate(train_reader()):
                 ts = time.time()
                 loss, acc, b_size = run_step(batch_id, data)
                 iters += 1
                 num_samples += len(data)
                 train_pass_acc.add(value=acc, weight=b_size)
+
+                duration = time.time() - ts
+                total_time += duration
+                count += len(data)
                 print(
                     "Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, "
-                    "Speed = %.2f img/s" % (pass_id, iters, loss, acc,
-                                            len(data) / (time.time() - ts))
+                    "Speed = %.2f (%.2f) img/s" % (pass_id, iters, loss, acc,
+                                                   len(data) / duration,
+                                                   count / total_time)
                 )  # The accuracy is the accumulation of batches, but not the current batch.
 
             pass_elapsed = time.time() - start_time
diff --git a/cmake/external/grpc.cmake b/cmake/external/grpc.cmake
index e90948782bb5e333bbdb47ef9d61c1e37e3cf9e4..ef520b128794d44df03fd11a916d8a791cba5c2a 100644
--- a/cmake/external/grpc.cmake
+++ b/cmake/external/grpc.cmake
@@ -33,7 +33,7 @@ ExternalProject_Add(
     extern_grpc
     DEPENDS protobuf zlib
     GIT_REPOSITORY "https://github.com/grpc/grpc.git"
-    GIT_TAG "v1.10.x"
+    GIT_TAG "v1.8.x"
     PREFIX          ${GRPC_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CONFIGURE_COMMAND ""
diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc
index 4e431561f81b2a84c06dff9fcb041317ebc84ae3..55be9b6c3bbceba34e8e9ba94e684d2536b7d25e 100644
--- a/paddle/fluid/framework/executor.cc
+++ b/paddle/fluid/framework/executor.cc
@@ -350,12 +350,9 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
       }
     }
   }
-  platform::DeviceContextPool::Instance().Get(place_)->Wait();
+  // platform::DeviceContextPool::Instance().Get(place_)->Wait();
   if (create_vars && create_local_scope) {
     scope->DeleteScope(local_scope);
-  } else {
-    // Delete the local scopes created in operators.
-    scope->DropKids();
   }
   if (FLAGS_benchmark) {
     VLOG(2) << "-------------------------------------------------------";
diff --git a/paddle/fluid/operators/detail/grpc_client.cc b/paddle/fluid/operators/detail/grpc_client.cc
index ae60ab15325ef101feb7270a4f5d840cb2112be0..47892b1bcc073d24ea617ea1c680138a88925177 100644
--- a/paddle/fluid/operators/detail/grpc_client.cc
+++ b/paddle/fluid/operators/detail/grpc_client.cc
@@ -19,6 +19,7 @@ limitations under the License. */
 #include <limits>
 
 #include "paddle/fluid/framework/threadpool.h"
+#include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace operators {
@@ -196,9 +197,14 @@ bool RPCClient::Wait() {
   const size_t kReqCnt = req_count_;
   bool a[kReqCnt];
   std::vector<std::future<void>> waits(req_count_);
+  std::mutex mu;
 
   for (int i = 0; i < req_count_; i++) {
-    waits[i] = framework::AsyncIO([i, &a, this] { a[i] = Proceed(); });
+    waits[i] = framework::AsyncIO([i, &a, &mu, this] {
+      bool ret = Proceed();
+      std::lock_guard<std::mutex> l(mu);
+      a[i] = ret;
+    });
   }
 
   for (int i = 0; i < req_count_; i++) {
diff --git a/paddle/fluid/operators/detail/grpc_server.cc b/paddle/fluid/operators/detail/grpc_server.cc
index eb114a47d99541402f748bfffcf6b10fde3e78e2..604321cd1f35abd4237e4b2ecc34dc148c4fd0d8 100644
--- a/paddle/fluid/operators/detail/grpc_server.cc
+++ b/paddle/fluid/operators/detail/grpc_server.cc
@@ -22,7 +22,10 @@ using ::grpc::ServerAsyncResponseWriter;
 namespace paddle {
 namespace operators {
 namespace detail {
-
+namespace {
+const int kNumHandleSendThreads = 20;
+const int kNumHandleGetThreads = 20;
+}  // namespace
 enum CallStatus { PROCESS = 0, FINISH };
 
 // reference:
@@ -63,18 +66,20 @@ class RequestSend final : public RequestBase {
   explicit RequestSend(GrpcService::AsyncService* service,
                        ::grpc::ServerCompletionQueue* cq, bool sync_mode,
                        framework::Scope* scope, ReceivedQueue* queue,
-                       const platform::DeviceContext* dev_ctx)
+                       const platform::DeviceContext* dev_ctx, int i)
       : RequestBase(service, cq, sync_mode, dev_ctx),
         queue_(queue),
-        responder_(&ctx_) {
+        responder_(&ctx_),
+        i_(i) {
     if (sync_mode_) {
       request_.reset(new VariableResponse(scope, dev_ctx_, false));
     } else {
       request_.reset(new VariableResponse(scope, dev_ctx_, true));
     }
     int method_id = static_cast<int>(detail::GrpcMethod::kSendVariable);
-    service_->RequestAsyncUnary(method_id, &ctx_, request_.get(), &responder_,
-                                cq_, cq_, this);
+    service_->RequestAsyncUnary(
+        method_id, &ctx_, request_.get(), &responder_, cq_, cq_,
+        reinterpret_cast<void*>(static_cast<intptr_t>(i)));
   }
 
   virtual ~RequestSend() {}
@@ -86,15 +91,17 @@ class RequestSend final : public RequestBase {
     VLOG(3) << "RequestSend " << var_name;
     queue_->Push(std::make_pair(var_name, request_));
 
-    sendrecv::VoidMessage reply;
-    responder_.Finish(reply, ::grpc::Status::OK, this);
     status_ = FINISH;
+    responder_.Finish(reply_, ::grpc::Status::OK,
+                      reinterpret_cast<void*>(static_cast<intptr_t>(i_)));
   }
 
  protected:
+  sendrecv::VoidMessage reply_;
   std::shared_ptr<VariableResponse> request_;
   ReceivedQueue* queue_;
   ServerAsyncResponseWriter<sendrecv::VoidMessage> responder_;
+  int i_;
 };
 
 class RequestGet final : public RequestBase {
@@ -103,14 +110,16 @@ class RequestGet final : public RequestBase {
                       ::grpc::ServerCompletionQueue* cq, bool sync_mode,
                       framework::Scope* scope,
                       const platform::DeviceContext* dev_ctx,
-                      framework::BlockingQueue<MessageWithName>* queue)
+                      framework::BlockingQueue<MessageWithName>* queue, int i)
       : RequestBase(service, cq, sync_mode, dev_ctx),
         responder_(&ctx_),
         scope_(scope),
-        queue_(queue) {
+        queue_(queue),
+        i_(i) {
     auto method_id = static_cast<int>(detail::GrpcMethod::kGetVariable);
-    service_->RequestAsyncUnary(method_id, &ctx_, &request_, &responder_, cq_,
-                                cq_, this);
+    service_->RequestAsyncUnary(
+        method_id, &ctx_, &request_, &responder_, cq_, cq_,
+        reinterpret_cast<void*>(static_cast<intptr_t>(i)));
   }
 
   virtual ~RequestGet() {}
@@ -123,13 +132,13 @@ class RequestGet final : public RequestBase {
     VLOG(3) << "RequestGet " << var_name;
     auto* var = scope_->FindVar(var_name);
 
-    ::grpc::ByteBuffer reply;
     if (var_name != FETCH_BARRIER_MESSAGE) {
-      SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply);
+      SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply_);
     }
 
-    responder_.Finish(reply, ::grpc::Status::OK, this);
     status_ = FINISH;
+    responder_.Finish(reply_, ::grpc::Status::OK,
+                      reinterpret_cast<void*>(static_cast<intptr_t>(i_)));
 
     if (var_name == FETCH_BARRIER_MESSAGE) {
       sendrecv::VariableMessage msg;
@@ -140,9 +149,11 @@ class RequestGet final : public RequestBase {
 
  protected:
   sendrecv::VariableMessage request_;
+  ::grpc::ByteBuffer reply_;
   ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_;
   framework::Scope* scope_;
   framework::BlockingQueue<MessageWithName>* queue_;
+  int i_;
 };
 
 class RequestPrefetch final : public RequestBase {
@@ -153,13 +164,15 @@ class RequestPrefetch final : public RequestBase {
                            const platform::DeviceContext* dev_ctx,
                            framework::Executor* executor,
                            framework::ProgramDesc* program,
-                           framework::ExecutorPrepareContext* prefetch_ctx)
+                           framework::ExecutorPrepareContext* prefetch_ctx,
+                           int i)
       : RequestBase(service, cq, sync_mode, dev_ctx),
         responder_(&ctx_),
         scope_(scope),
         executor_(executor),
         program_(program),
-        prefetch_ctx_(prefetch_ctx) {
+        prefetch_ctx_(prefetch_ctx),
+        i_(i) {
     if (sync_mode_) {
       request_.reset(new VariableResponse(scope, dev_ctx_, false));
     } else {
@@ -188,7 +201,8 @@ class RequestPrefetch final : public RequestBase {
 
     SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply);
 
-    responder_.Finish(reply, ::grpc::Status::OK, this);
+    responder_.Finish(reply, ::grpc::Status::OK,
+                      reinterpret_cast<void*>(static_cast<intptr_t>(i_)));
     status_ = FINISH;
   }
 
@@ -199,6 +213,7 @@ class RequestPrefetch final : public RequestBase {
   framework::Executor* executor_;
   framework::ProgramDesc* program_;
   framework::ExecutorPrepareContext* prefetch_ctx_;
+  int i_;
 };
 
 void AsyncGRPCServer::WaitClientGet(int count) {
@@ -232,20 +247,33 @@ void AsyncGRPCServer::RunSyncUpdate() {
   LOG(INFO) << "Server listening on " << address_
             << " selected port: " << selected_port_;
 
-  std::function<void()> send_register =
-      std::bind(&AsyncGRPCServer::TryToRegisterNewSendOne, this);
-  std::function<void()> get_register =
-      std::bind(&AsyncGRPCServer::TryToRegisterNewGetOne, this);
-  std::function<void()> prefetch_register =
-      std::bind(&AsyncGRPCServer::TryToRegisterNewPrefetchOne, this);
+  std::function<void(int)> send_register = std::bind(
+      &AsyncGRPCServer::TryToRegisterNewSendOne, this, std::placeholders::_1);
+  std::function<void(int)> get_register = std::bind(
+      &AsyncGRPCServer::TryToRegisterNewGetOne, this, std::placeholders::_1);
+  std::function<void(int)> prefetch_register =
+      std::bind(&AsyncGRPCServer::TryToRegisterNewPrefetchOne, this,
+                std::placeholders::_1);
+
+  for (int i = 0; i < kSendReqsBufSize; ++i) {
+    TryToRegisterNewSendOne(i);
+  }
+  for (int i = 0; i < kGetReqsBufSize; ++i) {
+    TryToRegisterNewGetOne(i);
+  }
+
+  for (int i = 0; i < kNumHandleSendThreads; ++i) {
+    t_sends_.emplace_back(
+        new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this,
+                                  cq_send_.get(), "cq_send", send_register)));
+  }
+  for (int i = 0; i < kNumHandleGetThreads; ++i) {
+    t_gets_.emplace_back(
+        new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this,
+                                  cq_get_.get(), "cq_get", get_register)));
+  }
 
   // TODO(wuyi): Run these "HandleRequest" in thread pool
-  t_send_.reset(
-      new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this,
-                                cq_send_.get(), "cq_send", send_register)));
-  t_get_.reset(
-      new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this,
-                                cq_get_.get(), "cq_get", get_register)));
   t_prefetch_.reset(new std::thread(
       std::bind(&AsyncGRPCServer::HandleRequest, this, cq_prefetch_.get(),
                 "cq_prefetch", prefetch_register)));
@@ -257,8 +285,27 @@ void AsyncGRPCServer::RunSyncUpdate() {
   condition_ready_.notify_all();
   // wait server
   server_->Wait();
-  t_send_->join();
-  t_get_->join();
+  for (int i = 0; i < kNumHandleSendThreads; ++i) {
+    t_sends_[i]->join();
+  }
+  for (int i = 0; i < kNumHandleGetThreads; ++i) {
+    t_gets_[i]->join();
+  }
+  {
+    std::lock_guard<std::mutex> l(cq_mutex_);
+    for (int i = 0; i < kSendReqsBufSize; ++i) {
+      if (send_reqs_[i]) {
+        delete send_reqs_[i];
+        send_reqs_[i] = nullptr;
+      }
+    }
+    for (int i = 0; i < kGetReqsBufSize; ++i) {
+      if (get_reqs_[i]) {
+        delete get_reqs_[i];
+        get_reqs_[i] = nullptr;
+      }
+    }
+  }
   t_prefetch_->join();
 }
 
@@ -276,47 +323,47 @@ void AsyncGRPCServer::ShutDown() {
   server_->Shutdown();
 }
 
-void AsyncGRPCServer::TryToRegisterNewSendOne() {
+void AsyncGRPCServer::TryToRegisterNewSendOne(int i) {
   std::unique_lock<std::mutex> lock(cq_mutex_);
   if (is_shut_down_) {
     VLOG(3) << "shutdown, do not TryToRegisterNewSendOne";
     return;
   }
   RequestSend* send = new RequestSend(&service_, cq_send_.get(), sync_mode_,
-                                      scope_, &var_recv_queue_, dev_ctx_);
+                                      scope_, &var_recv_queue_, dev_ctx_, i);
+  send_reqs_[i] = static_cast<RequestBase*>(send);
   VLOG(4) << "Create RequestSend status:" << send->Status();
 }
 
-void AsyncGRPCServer::TryToRegisterNewGetOne() {
+void AsyncGRPCServer::TryToRegisterNewGetOne(int i) {
   std::unique_lock<std::mutex> lock(cq_mutex_);
   if (is_shut_down_) {
     VLOG(3) << "shutdown, do not TryToRegisterNewGetOne";
     return;
   }
   RequestGet* get = new RequestGet(&service_, cq_get_.get(), sync_mode_, scope_,
-                                   dev_ctx_, &var_get_queue_);
+                                   dev_ctx_, &var_get_queue_, i);
+  get_reqs_[i] = static_cast<RequestBase*>(get);
   VLOG(4) << "Create RequestGet status:" << get->Status();
 }
 
-void AsyncGRPCServer::TryToRegisterNewPrefetchOne() {
+void AsyncGRPCServer::TryToRegisterNewPrefetchOne(int i) {
   std::unique_lock<std::mutex> lock(cq_mutex_);
   if (is_shut_down_) {
     VLOG(3) << "shutdown, do not TryToRegisterNewPrefetchOne";
     return;
   }
-  RequestPrefetch* prefetch =
-      new RequestPrefetch(&service_, cq_prefetch_.get(), sync_mode_, scope_,
-                          dev_ctx_, executor_, program_, prefetch_ctx_.get());
+  RequestPrefetch* prefetch = new RequestPrefetch(
+      &service_, cq_prefetch_.get(), sync_mode_, scope_, dev_ctx_, executor_,
+      program_, prefetch_ctx_.get(), i);
 
   VLOG(4) << "Create RequestPrefetch status:" << prefetch->Status();
 }
 
 // FIXME(typhoonzero): change cq_name to enum.
-void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq,
-                                    const std::string& cq_name,
-                                    std::function<void()> TryToRegisterNewOne) {
-  TryToRegisterNewOne();
-
+void AsyncGRPCServer::HandleRequest(
+    ::grpc::ServerCompletionQueue* cq, const std::string& cq_name,
+    std::function<void(int)> TryToRegisterNewOne) {
   void* tag = NULL;
   bool ok = false;
 
@@ -327,8 +374,7 @@ void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq,
       break;
     }
     VLOG(3) << "HandleRequest for " << cq_name << " get Next";
-
-    PADDLE_ENFORCE(tag);
+    int i = static_cast<int>(reinterpret_cast<intptr_t>(tag));
 
     if (sync_mode_) {
       // FIXME(typhoonzero): de-couple the barriers with recv_op
@@ -337,7 +383,17 @@ void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq,
       VLOG(3) << "HandleRequest for " << cq_name << " after WaitCond";
     }
 
-    RequestBase* base = reinterpret_cast<RequestBase*>(tag);
+    RequestBase* base = nullptr;
+    {
+      std::lock_guard<std::mutex> l(cq_mutex_);
+      if (cq_name == "cq_get") {
+        base = get_reqs_[i];
+      } else if (cq_name == "cq_send") {
+        base = send_reqs_[i];
+      } else {
+        CHECK(false);
+      }
+    }
     // reference:
     // https://github.com/tensorflow/tensorflow/issues/5596
     // https://groups.google.com/forum/#!topic/grpc-io/xftlRy-IQwM
@@ -345,19 +401,19 @@ void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq,
     if (!ok) {
       LOG(WARNING) << cq_name << " recv no regular event:argument name["
                    << base->GetReqName() << "]";
-      TryToRegisterNewOne();
+      TryToRegisterNewOne(i);
       delete base;
       continue;
     }
 
     switch (base->Status()) {
       case PROCESS: {
-        TryToRegisterNewOne();
         base->Process();
         VLOG(4) << cq_name << " PROCESS status:" << base->Status();
         break;
       }
       case FINISH: {
+        TryToRegisterNewOne(i);
         VLOG(4) << cq_name << " FINISH status:" << base->Status();
         delete base;
         break;
diff --git a/paddle/fluid/operators/detail/grpc_server.h b/paddle/fluid/operators/detail/grpc_server.h
index 238aaa29634a7eff65429c27aa3538a185723eb2..d70be1b7ce998a0faaaa256d62f6f1fcf5c267a0 100644
--- a/paddle/fluid/operators/detail/grpc_server.h
+++ b/paddle/fluid/operators/detail/grpc_server.h
@@ -17,6 +17,7 @@ limitations under the License. */
 #include <string>
 #include <thread>  // NOLINT
 #include <utility>
+#include <vector>
 
 #include "grpc++/grpc++.h"
 #include "paddle/fluid/framework/blocking_queue.h"
@@ -30,6 +31,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/detail/send_recv.grpc.pb.h"
 #include "paddle/fluid/operators/detail/send_recv.pb.h"
 #include "paddle/fluid/operators/detail/sendrecvop_utils.h"
+#include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace operators {
@@ -82,19 +84,25 @@ class AsyncGRPCServer final {
  protected:
   void HandleRequest(::grpc::ServerCompletionQueue *cq,
                      const std::string &cq_name,
-                     std::function<void()> TryToRegisterNewOne);
-  void TryToRegisterNewSendOne();
-  void TryToRegisterNewGetOne();
-  void TryToRegisterNewPrefetchOne();
+                     std::function<void(int)> TryToRegisterNewOne);
+  void TryToRegisterNewSendOne(int i);
+  void TryToRegisterNewGetOne(int i);
+  void TryToRegisterNewPrefetchOne(int i);
   void ShutdownQueue();
 
  private:
+  static const int kSendReqsBufSize = 100;
+  static const int kGetReqsBufSize = 100;
+
   std::mutex cq_mutex_;
   volatile bool is_shut_down_ = false;
   std::unique_ptr<::grpc::ServerCompletionQueue> cq_send_;
   std::unique_ptr<::grpc::ServerCompletionQueue> cq_get_;
   std::unique_ptr<::grpc::ServerCompletionQueue> cq_prefetch_;
 
+  RequestBase *send_reqs_[kSendReqsBufSize];
+  RequestBase *get_reqs_[kGetReqsBufSize];
+
   GrpcService::AsyncService service_;
   std::unique_ptr<::grpc::Server> server_;
 
@@ -113,8 +121,9 @@ class AsyncGRPCServer final {
   mutable int barrier_cond_step_;
   std::condition_variable barrier_condition_;
 
-  std::unique_ptr<std::thread> t_send_;
-  std::unique_ptr<std::thread> t_get_;
+  std::vector<std::unique_ptr<std::thread>> t_sends_;
+  std::vector<std::unique_ptr<std::thread>> t_gets_;
+
   std::unique_ptr<std::thread> t_prefetch_;
 
   std::unique_ptr<framework::ExecutorPrepareContext> prefetch_ctx_;
diff --git a/paddle/fluid/operators/detail/grpc_service.h b/paddle/fluid/operators/detail/grpc_service.h
index e6dab2f5a3a4280f3979417c3ca2d884a0b8ff2f..e0505c2b9d0903837713d7e0032b01ab091c2e04 100644
--- a/paddle/fluid/operators/detail/grpc_service.h
+++ b/paddle/fluid/operators/detail/grpc_service.h
@@ -25,6 +25,8 @@
 #include <grpc++/support/byte_buffer.h>
 #include "paddle/fluid/operators/detail/variable_response.h"
 
+#include "paddle/fluid/platform/profiler.h"
+
 // NOTE: This method was originally created by tensorflow
 //       (https://github.com/tensorflow/tensorflow/) we borrow this
 //       method and did some modifications so that we can parse gRPC
diff --git a/paddle/fluid/operators/detail/send_recv.proto b/paddle/fluid/operators/detail/send_recv.proto
index 9478c5702bcbf99fc88207b8c4843dbccf8a5925..078181909dff93c875bcf42ca54b3b77a3e87af2 100644
--- a/paddle/fluid/operators/detail/send_recv.proto
+++ b/paddle/fluid/operators/detail/send_recv.proto
@@ -73,7 +73,7 @@ message VariableMessage {
   // If true, the ps server will start profiling, the ps
   // server stops profiling and generates a profile to /tmp/profile_ps_*
   // when profile switches from true to false.
-  bool profile = 11;
+  int64 profile = 11;
 }
 
 message VoidMessage {}
diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.cc b/paddle/fluid/operators/detail/sendrecvop_utils.cc
index 07c43554bc6a0d71d688a5a5772d0ab3d2de319a..a9ea80c91736b78b969a985f4e89edd9e97f9c62 100644
--- a/paddle/fluid/operators/detail/sendrecvop_utils.cc
+++ b/paddle/fluid/operators/detail/sendrecvop_utils.cc
@@ -122,7 +122,13 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
   // 1 trainer returns true for ShouldSendProfileState(). It tells PS
   // servers the trainer's profiling state so that PS can follow the
   // trainer.
-  request.set_profile(platform::IsProfileEnabled());
+  if (platform::ShouldSendProfileState()) {
+    if (platform::IsProfileEnabled()) {
+      request.set_profile(1);
+    } else {
+      request.set_profile(2);
+    }
+  }
   if (!out_name.empty()) {
     request.set_out_varname(out_name);
   }
diff --git a/paddle/fluid/operators/detail/variable_response.cc b/paddle/fluid/operators/detail/variable_response.cc
index 462e303096e609c6797ca8cc16266ec3621623fc..2dfd9b2621e1d21e4bb02d9a2d50304638362e3e 100644
--- a/paddle/fluid/operators/detail/variable_response.cc
+++ b/paddle/fluid/operators/detail/variable_response.cc
@@ -449,8 +449,8 @@ int VariableResponse::Parse(Source* source) {
         break;
       }
       case sendrecv::VariableMessage::kProfileFieldNumber: {
-        bool profiling;
-        if (!input.ReadRaw(reinterpret_cast<void*>(&profiling), 1)) {
+        uint64_t profiling = 0;
+        if (!input.ReadVarint64(&profiling)) {
           return tag;
         }
         meta_.set_profile(profiling);
@@ -458,9 +458,9 @@ int VariableResponse::Parse(Source* source) {
         if (listener_id <= 0) {
           break;
         }
-        if (profiling && !platform::IsProfileEnabled()) {
+        if (profiling == 1 && !platform::IsProfileEnabled()) {
           platform::EnableProfiler(platform::ProfilerState::kCPU);
-        } else if (!profiling && platform::IsProfileEnabled()) {
+        } else if (profiling == 2 && platform::IsProfileEnabled()) {
           // TODO(panyx0718): Should we allow to customize file dir.
           platform::DisableProfiler(
               platform::EventSortingKey::kDefault,
diff --git a/paddle/fluid/platform/device_tracer.cc b/paddle/fluid/platform/device_tracer.cc
index c9e10631680a6ea3876f555a3a6e6c12f79b39d5..1a9be044e024e4b1dda5ef7d515c65f3a7513710 100644
--- a/paddle/fluid/platform/device_tracer.cc
+++ b/paddle/fluid/platform/device_tracer.cc
@@ -245,7 +245,6 @@ class DeviceTracerImpl : public DeviceTracer {
   void Enable() {
     std::lock_guard<std::mutex> l(trace_mu_);
     if (enabled_) {
-      fprintf(stderr, "DeviceTracer already enabled\n");
       return;
     }
     EnableActivity();