diff --git a/paddle/fluid/CMakeLists.txt b/paddle/fluid/CMakeLists.txt
index d274d96c29bdbf5973d568d783369c3975bdc436..e2e26fc5d40ed5584a14aacc5ba368138044f4a4 100644
--- a/paddle/fluid/CMakeLists.txt
+++ b/paddle/fluid/CMakeLists.txt
@@ -6,4 +6,4 @@ add_subdirectory(pybind)
 add_subdirectory(string)
 add_subdirectory(recordio)
 # NOTE: please add subdirectory inference at last.
-add_subdirectory(inference)
+#add_subdirectory(inference)
diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
index cdac00739bc48648b41751e644a953d0d310ffbf..18ec226801a7571b08efcdf672109281a1c83775 100644
--- a/paddle/fluid/framework/operator.cc
+++ b/paddle/fluid/framework/operator.cc
@@ -136,6 +136,12 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
     platform::SetDeviceId(dev_id);
 #endif
   }
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+  auto* dev_ctx = pool.Get(place);
+
+  // For profiling, don't move out of this function because that will result
+  // in the failure of multi-GPU profiling.
+  platform::RecordEvent record_event(Type(), dev_ctx);
   RunImpl(scope, place);
   VLOG(10) << "+ " << DebugStringEx(&scope);
 }
@@ -639,9 +645,6 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
   platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
   auto* dev_ctx = pool.Get(place);
 
-  // For profiling, don't move out of this function because that will result
-  // in the failure of multi-GPU profiling.
-  platform::RecordEvent record_event(Type(), dev_ctx);
   // check if op[type] has kernel registered.
   auto& all_op_kernels = AllOpKernels();
   auto kernels_iter = all_op_kernels.find(type_);
diff --git a/paddle/fluid/operators/feed_op.cc b/paddle/fluid/operators/feed_op.cc
index bcb3e63ed7dbc775c1de6c4522f0548ea48a6cf0..dc7ef664958238ddbd48745bd59cc7db28e49f5b 100644
--- a/paddle/fluid/operators/feed_op.cc
+++ b/paddle/fluid/operators/feed_op.cc
@@ -31,7 +31,6 @@ class FeedOp : public framework::OperatorBase {
                const platform::Place &place) const override {
     // get device context from pool
     auto *dev_ctx = platform::DeviceContextPool::Instance().Get(place);
-    platform::RecordEvent record_event(Type(), dev_ctx);
 
     auto feed_var_name = Input("X");
     auto *feed_var = scope.FindVar(feed_var_name);
diff --git a/paddle/fluid/operators/fetch_barrier_op.cc b/paddle/fluid/operators/fetch_barrier_op.cc
index 680fde19eefe57475b7526ebc29d4ff977a16977..d9cd956dfdff3d009d38ee5088f5396080580483 100644
--- a/paddle/fluid/operators/fetch_barrier_op.cc
+++ b/paddle/fluid/operators/fetch_barrier_op.cc
@@ -36,12 +36,6 @@ class FetchBarrierOp : public framework::OperatorBase {
   void RunImpl(const framework::Scope& scope,
                const platform::Place& place) const override {
     std::vector<std::string> eps = Attr<std::vector<std::string>>("endpoints");
-
-    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    auto& ctx = *pool.Get(place);
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
-
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
diff --git a/paddle/fluid/operators/fetch_op.cc b/paddle/fluid/operators/fetch_op.cc
index 1640a2a22c69a0e3ab81a2889d6105b2cf4162b7..c197b45e8196a47def6465128e8ca39d8daefed6 100644
--- a/paddle/fluid/operators/fetch_op.cc
+++ b/paddle/fluid/operators/fetch_op.cc
@@ -30,9 +30,6 @@ class FetchOp : public framework::OperatorBase {
  private:
   void RunImpl(const framework::Scope &scope,
                const platform::Place &place) const override {
-    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
-    platform::RecordEvent record_event(Type(), pool.Get(place));
-
     auto fetch_var_name = Input("X");
     auto *fetch_var = scope.FindVar(fetch_var_name);
     PADDLE_ENFORCE(fetch_var != nullptr,
diff --git a/paddle/fluid/operators/load_op.cc b/paddle/fluid/operators/load_op.cc
index ac35cf0b89bfaa0c0f8e64445f18a3bbd478e70a..27e26cb1b5c1e831f05dac299489628b92eaa58c 100644
--- a/paddle/fluid/operators/load_op.cc
+++ b/paddle/fluid/operators/load_op.cc
@@ -31,9 +31,6 @@ class LoadOp : public framework::OperatorBase {
  private:
   void RunImpl(const framework::Scope &scope,
                const platform::Place &place) const override {
-    auto *dev_ctx = platform::DeviceContextPool::Instance().Get(place);
-    platform::RecordEvent record_event(Type(), dev_ctx);
-
     // FIXME(yuyang18): We save variable to local file now, but we should change
     // it to save an output stream.
     auto filename = Attr<std::string>("file_path");
diff --git a/paddle/fluid/operators/recv_op.cc b/paddle/fluid/operators/recv_op.cc
index 1ba684014904e61a86bebacd7d29d7e10d313092..4a6ce938a5f337d035b21f562d46daf606236db0 100644
--- a/paddle/fluid/operators/recv_op.cc
+++ b/paddle/fluid/operators/recv_op.cc
@@ -40,8 +40,6 @@ class RecvOp : public framework::OperatorBase {
 
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
 
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>();
diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc
index d7f8e994afd7e656bd5a9dd7c5ab45f0d52fe88b..1866a86048acbefadcb4d82cd6309cd16f0352d6 100644
--- a/paddle/fluid/operators/send_barrier_op.cc
+++ b/paddle/fluid/operators/send_barrier_op.cc
@@ -39,11 +39,6 @@ class SendBarrierOp : public framework::OperatorBase {
     std::vector<std::string> eps = Attr<std::vector<std::string>>("endpoints");
     bool sync_mode = Attr<bool>("sync_mode");
 
-    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    auto& ctx = *pool.Get(place);
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
-
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc
index 829f310d4233c01a7fbb9ccf7427f6e47ce8d384..3cd42f2d059532b7090e66ce21de8e5cb014adf1 100644
--- a/paddle/fluid/operators/send_op.cc
+++ b/paddle/fluid/operators/send_op.cc
@@ -42,9 +42,6 @@ class SendOp : public framework::OperatorBase {
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
 
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
-
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>();