diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt
index c701a2ad63048f69e8443fee7434928698aa20cc..b419c8c292a7afe07d1189cab15c1a3892bd0dfe 100644
--- a/paddle/fluid/framework/CMakeLists.txt
+++ b/paddle/fluid/framework/CMakeLists.txt
@@ -177,7 +177,7 @@ else()
 endif()
 
 cc_library(parallel_executor SRCS parallel_executor.cc DEPS
-        threaded_ssa_graph_executor scope_buffered_ssa_graph_executor
+        threaded_ssa_graph_executor scope_buffered_ssa_graph_executor parallel_ssa_graph_executor
         graph build_strategy
         fast_threaded_ssa_graph_executor variable_helper)
 
diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt
index 93288936fea1fae897dc26e6d8850da612960333..65247533228cda6cfd9bcaf85734df4901ba6c50 100644
--- a/paddle/fluid/framework/details/CMakeLists.txt
+++ b/paddle/fluid/framework/details/CMakeLists.txt
@@ -54,6 +54,8 @@ cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUT
 cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
         simple_threadpool device_context)
 
+cc_library(parallel_ssa_graph_executor SRCS parallel_ssa_graph_executor.cc DEPS threaded_ssa_graph_executor)
+
 cc_test(broadcast_op_test SRCS broadcast_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
         device_context broadcast_op_handle)
 cc_test(gather_op_test SRCS gather_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
diff --git a/paddle/fluid/framework/details/all_reduce_op_handle.cc b/paddle/fluid/framework/details/all_reduce_op_handle.cc
index e8bf53e160e7382122c3c2f92a152fea058a032e..ae17ea8a1541a2f56d3509d1af353ef82fe38eba 100644
--- a/paddle/fluid/framework/details/all_reduce_op_handle.cc
+++ b/paddle/fluid/framework/details/all_reduce_op_handle.cc
@@ -46,20 +46,27 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
 #endif
 
 void AllReduceOpHandle::RunImpl() {
+  int64_t start_ts = GetTS();
+  int64_t func_ts = GetTS();
+  VLOG(5) << "all_reduce_op_handle::RunImpl start";
   platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
 
 // FIXME(typhoonzero): If scope0(global scope) have NCCL_ID_VAR,
 // this is a distributed or inter-process call, find a better way.
 #ifdef PADDLE_WITH_CUDA
   if (NoDummyInputSize() == 1 &&
-      local_scopes_[0]->FindLocalVar(NCCL_ID_VARNAME) == nullptr) {
+      local_scopes_[0]->FindVar(NCCL_ID_VARNAME) == nullptr) {
 #else
   if (NoDummyInputSize() == 1) {
 #endif
     return;  // No need to all reduce when GPU count = 1;
   } else {
     // Wait input done
+    start_ts = GetTS();
     WaitInputVarGenerated();
+    VLOG(5) << "all_reduce_op_handle wait input var spent: "
+            << GetTS() - start_ts << " (ns).";
+    start_ts = GetTS();
     auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
     auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
     PADDLE_ENFORCE_EQ(
@@ -100,6 +107,8 @@ void AllReduceOpHandle::RunImpl() {
         }
 
         int dev_id = boost::get<platform::CUDAPlace>(p).device;
+        VLOG(5) << "call allreduce: " << in_var_handles[i]->name_
+                << " on dev: " << dev_id;
         auto &nccl_ctx = nccl_ctxs_->at(dev_id);
         auto stream = nccl_ctx.stream();
         auto comm = nccl_ctx.comm_;
@@ -110,11 +119,20 @@ void AllReduceOpHandle::RunImpl() {
         });
       }
       this->RunAndRecordEvent([&] {
-        platform::NCCLGroupGuard guard;
-        for (auto &call : all_reduce_calls) {
-          call();
+        // TODO(Yancey1989): need allreduce operator to avoid this flag
+        if (nccl_ctxs_->need_group_call_) {
+          platform::NCCLGroupGuard guard;
+          for (auto &call : all_reduce_calls) {
+            call();
+          }
+        } else {
+          // only used in executor_type == ParallalGraph, one thread one GPU
+          // TODO(Yancey1989): use allreduce operator to avoid this tricky.
+          PADDLE_ENFORCE(all_reduce_calls.size() == 1UL);
+          all_reduce_calls[0]();
         }
       });
+
 #else
       PADDLE_THROW("Not compiled with CUDA");
 #endif
@@ -144,6 +162,8 @@ void AllReduceOpHandle::RunImpl() {
       }
     }
   }
+  VLOG(5) << "all_reduce_op_handle Impl spent: " << GetTS() - func_ts
+          << " (ns).";
 }
 
 std::string AllReduceOpHandle::Name() const { return "all_reduce"; }
diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc
index 1e1b945f63cf480308c05ffc0f9a3b9f0b0da02b..04c1061536b778fa44e67504cd22846931e039c9 100644
--- a/paddle/fluid/framework/details/build_strategy.cc
+++ b/paddle/fluid/framework/details/build_strategy.cc
@@ -118,6 +118,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
   std::unique_ptr<ir::Graph> graph(new ir::Graph(main_program));
 
   for (std::shared_ptr<ir::Pass> &pass : pass_builder_->AllPasses()) {
+    VLOG(5) << "run pass: " << pass->Type();
     if (pass->Type() == "multi_devices_pass") {
       pass->Erase("places");
       pass->SetNotOwned<const std::vector<platform::Place>>("places", &places);
diff --git a/paddle/fluid/framework/details/computation_op_handle.cc b/paddle/fluid/framework/details/computation_op_handle.cc
index 7ad1e40c600c6e70cea822fac777ff20163078e6..35ba99a879396555dbbf11c4fe37d40187f40c89 100644
--- a/paddle/fluid/framework/details/computation_op_handle.cc
+++ b/paddle/fluid/framework/details/computation_op_handle.cc
@@ -33,10 +33,18 @@ void ComputationOpHandle::RunImpl() {
     op_->Run(*scope_->FindVar(kLocalExecScopeName)->Get<Scope *>(), place_);
   };
 
-  if (is_lock_and_record_event_free_) {
+  if (Name().compare("conv2d") || Name().compare("conv2d_grad")) {
+    int64_t start_ts = GetTS();
+    auto varname = DynamicCast<VarHandle>(this->Outputs())[0]->name_;
     run_func();
+    VLOG(5) << Name() << "_op_handle: " << varname
+            << " spent: " << GetTS() - start_ts << " (ns).";
   } else {
-    this->RunAndRecordEvent(run_func);
+    if (is_lock_and_record_event_free_) {
+      run_func();
+    } else {
+      this->RunAndRecordEvent(run_func);
+    }
   }
 }
 
diff --git a/paddle/fluid/framework/details/computation_op_handle.h b/paddle/fluid/framework/details/computation_op_handle.h
index 662a91d6b4dfcfed563fdf2e46c22f83f90b40af..5346b56dd63d19e3cffedf2e28bba3942abb3856 100644
--- a/paddle/fluid/framework/details/computation_op_handle.h
+++ b/paddle/fluid/framework/details/computation_op_handle.h
@@ -17,6 +17,7 @@
 #include <string>
 #include <vector>
 
+#include "paddle/fluid/framework/details/container_cast.h"
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
diff --git a/paddle/fluid/framework/details/execution_strategy.h b/paddle/fluid/framework/details/execution_strategy.h
index 15c496130c2b6c7643ff96661be09e5ac4870344..d3d5b6bf541597c6c217dc8d866ec44d1244a132 100644
--- a/paddle/fluid/framework/details/execution_strategy.h
+++ b/paddle/fluid/framework/details/execution_strategy.h
@@ -20,7 +20,7 @@ namespace framework {
 namespace details {
 
 struct ExecutionStrategy {
-  enum ExecutorType { kDefault = 0, kExperimental = 1 };
+  enum ExecutorType { kDefault = 0, kExperimental = 1, kParallelGraph = 2 };
 
   size_t num_threads_{0};
   bool use_cuda_{true};
diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc
index cbae5321d9a3fdcaebca2bd9111fff6a10b235a8..1bd238357a7c22ef42cfbefb13837e693668d19a 100644
--- a/paddle/fluid/framework/details/multi_devices_graph_pass.cc
+++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc
@@ -300,7 +300,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
   auto nodes = graph->ReleaseNodes();
   ir::Graph &result = *graph;
 
-  int num_trainers = Get<int>(kNumTrainers);
+  // int num_trainers = Get<int>(kNumTrainers);
 
   for (auto &node : nodes) {
     if (node->IsVar() && node->Var()) {
@@ -329,6 +329,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
   std::unordered_map<std::string, int> sharded_var_device;
 
   for (ir::Node *node : sorted_ops) {
+    VLOG(5) << "op name: " << node->Op()->Type();
     if (boost::get<int>(
             node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) ==
         static_cast<int>(OpRole::kRPC)) {
@@ -365,9 +366,11 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
       // is true only for the op that scale the final scalar loss.
       // It also assumes backward op will always follow the forward op in
       // the block.
+      VLOG(5) << "this is loss scale op!";
       is_forwarding = false;
     } else {
       int op_dev_id = GetOpDeviceID(result, node, sharded_var_device);
+      VLOG(5) << "on device id: " << op_dev_id;
       if (op_dev_id != -1) {  // This op only runs on one specific device.
         CreateComputationalOp(&result, node, op_dev_id);
         for (ir::Node *n : node->outputs) {
@@ -386,7 +389,8 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
           CreateComputationalOps(&result, node, places_.size());
         }
 
-        if (!is_forwarding && (places_.size() > 1 || num_trainers > 1)) {
+        // if (!is_forwarding && (places_.size() > 1 || num_trainers > 1)) {
+        if (!is_forwarding && nccl_ctxs_->contexts_.size() > 1) {
           // Currently, we assume that once gradient is generated, it can be
           // broadcast, and each gradient is only broadcast once.
           if (static_cast<bool>(boost::get<int>(node->Op()->GetAttr(
diff --git a/paddle/fluid/framework/details/op_handle_base.cc b/paddle/fluid/framework/details/op_handle_base.cc
index 4822627ac3b65972f41d9a23d9fe3dba3de3f97d..d68d1ce71d0fcc6856f078164a3fd235aad236f7 100644
--- a/paddle/fluid/framework/details/op_handle_base.cc
+++ b/paddle/fluid/framework/details/op_handle_base.cc
@@ -41,6 +41,7 @@ OpHandleBase::~OpHandleBase() {
 
 void OpHandleBase::Run(bool use_cuda) {
 #ifdef PADDLE_WITH_CUDA
+  int64_t start_ts = 0;
   if (events_.empty() && use_cuda) {
     for (auto &p : dev_ctxes_) {
       int dev_id = boost::get<platform::CUDAPlace>(p.first).device;
@@ -52,7 +53,6 @@ void OpHandleBase::Run(bool use_cuda) {
 #else
   PADDLE_ENFORCE(!use_cuda);
 #endif
-
   RunImpl();
 }
 
@@ -125,6 +125,7 @@ bool OpHandleBase::NeedWait(VarHandleBase *in_var) {
 void OpHandleBase::RunAndRecordEvent(const std::function<void()> &callback) {
 #ifdef PADDLE_WITH_CUDA
   if (!events_.empty()) {  // Use event
+    VLOG(5) << "events not empty";
     std::function<void()> method = callback;
     for (auto &p : dev_ctxes_) {
       method = [method, p, this]() {
diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h
index ba12ca3c61c05b3e856fffa8353d4ec5bf79bc39..88c78e0678c3ef547873d67612bdfe4e037722a9 100644
--- a/paddle/fluid/framework/details/op_handle_base.h
+++ b/paddle/fluid/framework/details/op_handle_base.h
@@ -26,7 +26,6 @@ namespace framework {
 namespace details {
 
 constexpr char kLocalExecScopeName[] = "@LCOAL_SCOPE@";
-
 // Wraps ir::Node and provide helper utilities.
 // It's responsible for populating necessary fields of ir::Node.
 class OpHandleBase {
diff --git a/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc b/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
new file mode 100644
index 0000000000000000000000000000000000000000..72beb74aa4815dbaa092dcbdc749a30bd3eb549d
--- /dev/null
+++ b/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
@@ -0,0 +1,66 @@
+//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+ParallelSSAGraphExecutor::ParallelSSAGraphExecutor(
+    const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
+    const std::vector<platform::Place> &places,
+    std::vector<std::unique_ptr<ir::Graph>> graphs)
+    : strategy_(std::move(strategy)),
+      local_scopes_(std::move(local_scopes)),
+      places_(std::move(places)),
+      graphs_(std::move(graphs)),
+      pool_(places.size() >= 2 ? new ::ThreadPool(places.size()) : nullptr) {
+  PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
+  for (size_t i = 0; i < places.size(); ++i) {
+    std::vector<framework::Scope *> scopes = {local_scopes_[i]};
+    std::vector<platform::Place> places = {places_[i]};
+    executors_.emplace_back(new details::ThreadedSSAGraphExecutor(
+        strategy_, scopes, places, std::move(graphs_[i])));
+  }
+}
+
+FeedFetchList ParallelSSAGraphExecutor::Run(
+    const std::vector<std::string> &fetch_tensors) {
+  std::vector<std::future<void>> run_futures;
+  FeedFetchList fetch_data;
+
+  for (size_t i = 0; i < places_.size(); ++i) {
+    auto call = [this, i] {
+      // FIXME(Yancey1989): need to fix fetch data failed.
+      std::vector<std::string> empty;
+      executors_[i]->Run(empty);
+    };
+    if (pool_) {
+      run_futures.emplace_back(pool_->enqueue(std::move(call)));
+    } else {
+      call();
+    }
+  }
+  if (pool_) {
+    for (auto &f : run_futures) {
+      f.wait();
+    }
+  }
+  return fetch_data;
+}
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle
diff --git a/paddle/fluid/framework/details/parallel_ssa_graph_executor.h b/paddle/fluid/framework/details/parallel_ssa_graph_executor.h
new file mode 100644
index 0000000000000000000000000000000000000000..c0ba1577f7ffa7924114104ebede16021e7f540d
--- /dev/null
+++ b/paddle/fluid/framework/details/parallel_ssa_graph_executor.h
@@ -0,0 +1,51 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <string>
+#include <vector>
+
+#include "ThreadPool.h"
+#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
+#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+class ParallelSSAGraphExecutor : public SSAGraphExecutor {
+ public:
+  ParallelSSAGraphExecutor(const ExecutionStrategy &strategy,
+                           const std::vector<Scope *> &local_scopes,
+                           const std::vector<platform::Place> &places,
+                           std::vector<std::unique_ptr<ir::Graph>> graphs);
+  ~ParallelSSAGraphExecutor() final = default;
+  const ir::Graph &Graph() const override { return *graphs_[0]; }
+
+  FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
+
+ private:
+  ExecutionStrategy strategy_;
+  std::vector<Scope *> local_scopes_;
+  std::vector<platform::Place> places_;
+  std::vector<std::unique_ptr<ir::Graph>> graphs_;
+  std::unique_ptr<::ThreadPool> pool_;
+
+  std::vector<std::unique_ptr<details::ThreadedSSAGraphExecutor>> executors_;
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle
diff --git a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc
index 499246a9856bb3ba67a155c6f00c3ad06af50edf..abc6b9f559ead54537c1f6f5d39811366aba88fd 100644
--- a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc
+++ b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc
@@ -27,39 +27,40 @@ namespace framework {
 namespace details {
 ScopeBufferedSSAGraphExecutor::ScopeBufferedSSAGraphExecutor(
     ExecutionStrategy strategy, std::vector<Scope *> local_scopes,
-    std::vector<VariableInfo> var_infos, std::vector<platform::Place> places,
+    std::vector<std::vector<VariableInfo>> var_infos_list,
+    std::vector<platform::Place> places,
     std::unique_ptr<SSAGraphExecutor> &&underlying_executor)
     : strategy_(std::move(strategy)),
       underlying_executor_(std::move(underlying_executor)),
       local_scopes_(std::move(local_scopes)),
-      var_infos_(std::move(var_infos)),
+      var_infos_list_(std::move(var_infos_list)),
       places_(std::move(places)) {}
 
 FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
     const std::vector<std::string> &fetch_tensors) {
   if (drop_scope_counter_ == 0) {
     // Create local scopes.
-    for (auto it = local_scopes_.rbegin(); it != local_scopes_.rend(); ++it) {
-      auto &scope = *it;
+    for (size_t i = 0; i < local_scopes_.size(); ++i) {
+      auto &scope = local_scopes_[i];
       Scope &local_scope = scope->NewScope();
       *scope->Var(details::kLocalExecScopeName)->GetMutable<Scope *>() =
           &local_scope;
-
-      for (auto &info : var_infos_) {
-        if (scope->FindVar(info.name_) != nullptr) {
-          continue;
-        }
-
-        if (info.persistable_) {  // Persistable
-          InitializeVariable(scope->Var(info.name_), info.type_);
-        } else {
-          InitializeVariable(local_scope.Var(info.name_), info.type_);
+      for (auto &var_infos : var_infos_list_) {
+        for (auto &info : var_infos) {
+          if (scope->FindVar(info.name_) != nullptr) {
+            continue;
+          }
+          if (info.persistable_) {  // Persistable
+            InitializeVariable(scope->Var(info.name_), info.type_);
+          } else {
+            InitializeVariable(local_scope.Var(info.name_), info.type_);
+          }
         }
       }
     }
   }
   std::vector<framework::LoDTensor> fetch_data;
-  std::exception_ptr eptr;
+  std::exception_ptr eptr = nullptr;
   try {
     fetch_data = underlying_executor_->Run(fetch_tensors);
   } catch (...) {
@@ -71,9 +72,13 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
 
 #ifdef PADDLE_WITH_CUDA
   const std::string gc_name = "garbage_collector";
-  DeviceGarbageCollectorMap *gc =
-      Graph().Has(gc_name) ? &(Graph().Get<DeviceGarbageCollectorMap>(gc_name))
-                           : nullptr;
+  DeviceGarbageCollectorMap *gc = nullptr;
+  // FIXME(Yancey1989): need to fix gc failed on parallel graph mode
+  if (strategy_.type_ != ExecutionStrategy::kParallelGraph) {
+    gc = Graph().Has(gc_name)
+             ? &(Graph().Get<DeviceGarbageCollectorMap>(gc_name))
+             : nullptr;
+  }
 #endif
 
   if (!fetch_tensors.empty() ||
diff --git a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h
index 5e87e0bf50b51d2b630aba06a5907dd721754d1f..51230d4a42abd8efa8dc395db2117405cdb90ffc 100644
--- a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h
+++ b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h
@@ -38,7 +38,8 @@ class ScopeBufferedSSAGraphExecutor : public SSAGraphExecutor {
  public:
   ScopeBufferedSSAGraphExecutor(
       ExecutionStrategy strategy, std::vector<Scope*> local_scopes,
-      std::vector<VariableInfo> var_infos, std::vector<platform::Place> places,
+      std::vector<std::vector<VariableInfo>> var_info_list,
+      std::vector<platform::Place> places,
       std::unique_ptr<SSAGraphExecutor>&& underlying_executor);
 
   const ir::Graph& Graph() const override {
@@ -53,7 +54,7 @@ class ScopeBufferedSSAGraphExecutor : public SSAGraphExecutor {
   ExecutionStrategy strategy_;
   std::unique_ptr<SSAGraphExecutor> underlying_executor_;
   std::vector<Scope*> local_scopes_;
-  std::vector<VariableInfo> var_infos_;
+  std::vector<std::vector<VariableInfo>> var_infos_list_;
   std::vector<platform::Place> places_;
 };
 }  // namespace details
diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h
index 24da56c09e3e0f3894d58e5af8838c98e3e1e67c..b45afbc0461d212095f422a71e0eee27572a2f39 100644
--- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h
+++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h
@@ -24,6 +24,7 @@
 #include <functional>
 #include "ThreadPool.h"  // ThreadPool in thrird party
 #include "paddle/fluid/framework/blocking_queue.h"
+#include "paddle/fluid/framework/details/computation_op_handle.h"
 #include "paddle/fluid/framework/details/exception_holder.h"
 #include "paddle/fluid/framework/details/execution_strategy.h"
 #include "paddle/fluid/framework/details/fetch_op_handle.h"
diff --git a/paddle/fluid/framework/details/var_handle.cc b/paddle/fluid/framework/details/var_handle.cc
index 30da029ca2a90e7faa6288557ff2f1aeb21cc1c6..7de6025a28a1992786b73d53d456984e0cf418c5 100644
--- a/paddle/fluid/framework/details/var_handle.cc
+++ b/paddle/fluid/framework/details/var_handle.cc
@@ -20,7 +20,7 @@ namespace details {
 
 VarHandleBase::~VarHandleBase() {}
 
-VarHandle::~VarHandle() { VLOG(4) << "deleting var handle " << DebugString(); }
+VarHandle::~VarHandle() { VLOG(5) << "deleting var handle " << DebugString(); }
 
 std::string VarHandle::DebugString() const {
   std::stringstream ss;
diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc
index b98408ee7726768a108772329b8dc95c2df3c891..ff3d76fb01cd8e410a712c42a227e441aeede396 100644
--- a/paddle/fluid/framework/parallel_executor.cc
+++ b/paddle/fluid/framework/parallel_executor.cc
@@ -26,6 +26,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h"
 #include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
 #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -53,6 +54,7 @@ class ParallelExecutorPrivate {
   std::vector<Scope *> local_scopes_;
   Scope *global_scope_;  // not owned
   std::unique_ptr<details::SSAGraphExecutor> executor_;
+  std::vector<std::unique_ptr<details::SSAGraphExecutor>> executors_;
 
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
@@ -84,6 +86,9 @@ ParallelExecutor::ParallelExecutor(
     PADDLE_ENFORCE(places.size() > 1,
                    "If you set build_strategy.reduce with 'Reduce',"
                    "the number of places must be greater than 1.");
+    PADDLE_ENFORCE(exec_strategy.type_ != ExecutionStrategy::kParallelGraph,
+                   "You should set build_strategy.reduce with 'AllReduce' for "
+                   "ParallelGraph executor type");
   }
 
   // Step 1. Bcast the params to devs.
@@ -106,31 +111,55 @@ ParallelExecutor::ParallelExecutor(
 // Bcast Parameters to all GPUs
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME);
-    ncclUniqueId *nccl_id = nullptr;
+    std::unique_ptr<ncclUniqueId> nccl_id = nullptr;
+    bool need_group_call = true;
     if (nccl_id_var != nullptr) {
-      nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
+      nccl_id.reset(nccl_id_var->GetMutable<ncclUniqueId>());
+    } else if (exec_strategy.type_ == ExecutionStrategy::kParallelGraph) {
+      nccl_id.reset(new ncclUniqueId());
+      PADDLE_ENFORCE(platform::dynload::ncclGetUniqueId(nccl_id.get()));
+      *member_->global_scope_->Var(NCCL_ID_VARNAME)
+           ->GetMutable<ncclUniqueId>() = *nccl_id.get();
+      need_group_call = false;
+    } else {
+      // init nccl_id in NCCLContextMap
     }
+
     member_->nccl_ctxs_.reset(new platform::NCCLContextMap(
-        member_->places_, nccl_id, num_trainers, trainer_id));
+        member_->places_, nccl_id.get(), num_trainers, trainer_id,
+        need_group_call));
 #else
     PADDLE_THROW("Not compiled with CUDA");
 #endif
   }
-
   if (member_->local_scopes_.size() != 1 && local_scopes.empty()) {
     BCastParamsToDevices(bcast_vars);
   }
-// Startup Program has been run. All local scopes has correct parameters.
+  // Startup Program has been run. All local scopes has correct parameters.
 
-// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
-// ncclOp
+  // Step 2. Convert main_program to SSA form and dependency graph. Also, insert
+  // ncclOp
+  std::vector<std::unique_ptr<ir::Graph>> graphs;
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-  std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
-      main_program, member_->places_, loss_var_name, params,
-      member_->local_scopes_, member_->use_cuda_, member_->nccl_ctxs_.get());
+  if (exec_strategy.type_ == ExecutionStrategy::kParallelGraph) {
+    for (size_t i = 0; i < member_->places_.size(); ++i) {
+      std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
+          main_program, {member_->places_[i]}, loss_var_name, params,
+          {member_->local_scopes_[i]}, member_->use_cuda_,
+          member_->nccl_ctxs_.get());
+      graphs.push_back(std::move(graph));
+    }
+  } else {
+    std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
+        main_program, member_->places_, loss_var_name, params,
+        member_->local_scopes_, member_->use_cuda_, member_->nccl_ctxs_.get());
+    graphs.push_back(std::move(graph));
+  }
 
   auto max_memory_size = GetEagerDeletionThreshold();
-  if (max_memory_size >= 0) {
+  // FIXME(Yancey1989): need to fix on parallel graph mode
+  if (max_memory_size >= 0 &&
+      exec_strategy.type_ != ExecutionStrategy::kParallelGraph) {
     for (auto &place : member_->places_) {
       if (!platform::is_gpu_place(place)) continue;
       auto gpu_place = boost::get<platform::CUDAPlace>(place);
@@ -143,40 +172,48 @@ ParallelExecutor::ParallelExecutor(
       }
     }
     if (!gcs_.empty()) {
-      auto ref_cnt_pass =
-          ir::PassRegistry::Instance().Get("reference_count_pass");
-      ref_cnt_pass->SetNotOwned(details::kGlobalReferenceCount, &ref_cnts_);
-      ref_cnt_pass->SetNotOwned(details::kCurReferenceCount, &cur_ref_cnts_);
-      ref_cnt_pass->SetNotOwned(details::kGarbageCollector, &gcs_);
-      graph = ref_cnt_pass->Apply(std::move(graph));
-      graph->SetNotOwned("garbage_collector", &gcs_);
+      for (size_t i = 0; i < graphs.size(); ++i) {
+        auto ref_cnt_pass =
+            ir::PassRegistry::Instance().Get("reference_count_pass");
+        ref_cnt_pass->SetNotOwned(details::kGlobalReferenceCount, &ref_cnts_);
+        ref_cnt_pass->SetNotOwned(details::kCurReferenceCount, &cur_ref_cnts_);
+        ref_cnt_pass->SetNotOwned(details::kGarbageCollector, &gcs_);
+        graphs[0] = ref_cnt_pass->Apply(std::move(graphs[i]));
+        graphs[0]->SetNotOwned("garbage_collector", &gcs_);
+      }
     }
   }
 #else
   std::unique_ptr<ir::Graph> graph =
       build_strategy.Apply(main_program, member_->places_, loss_var_name,
                            params, member_->local_scopes_, member_->use_cuda_);
+  graphs.push_back(std::move(graph));
 #endif
 
   // Step 3. Create vars in each scope. Passes may also create new vars.
   //         skip control vars and empty vars
-  std::vector<details::VariableInfo> var_infos;
-  for (auto &node : graph->Nodes()) {
-    if (node->IsVar() && !node->IsCtrlVar() && node->Var()) {
-      var_infos.emplace_back();
-      var_infos.back().name_ = node->Var()->Name();
-      var_infos.back().type_ = node->Var()->GetType();
-      var_infos.back().persistable_ = node->Var()->Persistable();
+  std::vector<std::vector<details::VariableInfo>> var_infos_list;
+  for (size_t i = 0; i < graphs.size(); ++i) {
+    std::vector<details::VariableInfo> var_infos;
+    for (auto &node : graphs[i]->Nodes()) {
+      if (node->IsVar() && !node->IsCtrlVar() && node->Var()) {
+        var_infos.emplace_back();
+        var_infos.back().name_ = node->Var()->Name();
+        var_infos.back().type_ = node->Var()->GetType();
+        var_infos.back().persistable_ = node->Var()->Persistable();
+      }
     }
+    var_infos_list.emplace_back(std::move(var_infos));
   }
+
   // If the loss_var_name is given, the number of graph should be only one.
   if (loss_var_name.size()) {
-    size_t graph_num = ir::GraphNum(*graph);
+    size_t graph_num = ir::GraphNum(*graphs[0]);
     if (graph_num > 1) {
       LOG(WARNING)
           << "The number of graph should be only one, "
              "but the current graph has "
-          << ir::GraphNum(*graph)
+          << ir::GraphNum(*graphs[0])
           << " sub_graphs. If you want to see the nodes of the "
              "sub_graphs, you should use 'FLAGS_print_sub_graph_dir' "
              "to specify the output dir. NOTES: if you not do training, "
@@ -185,15 +222,42 @@ ParallelExecutor::ParallelExecutor(
   }
 
   if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
+    /**
+    for (size_t i = 0; i < member_->places_.size(); ++i) {
+      std::vector<details::VariableInfo> var_infos;
+      for (auto &node : graphs[i]->Nodes()) {
+        if (node->IsVar() && !node->IsCtrlVar() && node->Var()) {
+          var_infos.emplace_back();
+          var_infos.back().name_ = node->Var()->Name();
+          var_infos.back().type_ = node->Var()->GetType();
+          var_infos.back().persistable_ = node->Var()->Persistable();
+        }
+      }
+
+      std::vector<platform::Place> places = {member_->places_[i]};
+      std::vector<framework::Scope *> scopes = {member_->local_scopes_[i]};
+      std::unique_ptr<details::ThreadedSSAGraphExecutor> p(new
+    details::ThreadedSSAGraphExecutor(
+        exec_strategy, scopes, places, std::move(graphs[i])));
+
+      member_->executors_.push_back(std::move(p));
+
+      member_->executors_[i].reset(new details::ScopeBufferedSSAGraphExecutor(
+        exec_strategy, scopes, std::move(var_infos), places,
+        std::move(member_->executors_[i])));
+    }**/
     member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
-        exec_strategy, member_->local_scopes_, places, std::move(graph)));
+        exec_strategy, member_->local_scopes_, places, std::move(graphs[0])));
+  } else if (exec_strategy.type_ == ExecutionStrategy::kParallelGraph) {
+    member_->executor_.reset(new details::ParallelSSAGraphExecutor(
+        exec_strategy, member_->local_scopes_, places, graphs));
   } else {
     member_->executor_.reset(new details::FastThreadedSSAGraphExecutor(
-        exec_strategy, member_->local_scopes_, places, std::move(graph)));
+        exec_strategy, member_->local_scopes_, places, std::move(graphs[0])));
   }
 
   member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor(
-      exec_strategy, member_->local_scopes_, std::move(var_infos),
+      exec_strategy, member_->local_scopes_, std::move(var_infos_list),
       member_->places_, std::move(member_->executor_)));
 }
 
diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h
index ef09b98b2aa91a9d729b94d15dbb676dde4092b6..319701f1eb8784202bd94be559c3802784499d33 100644
--- a/paddle/fluid/framework/parallel_executor.h
+++ b/paddle/fluid/framework/parallel_executor.h
@@ -20,6 +20,8 @@ limitations under the License. */
 #include <unordered_set>
 #include <vector>
 
+#include "ThreadPool.h"
+
 #include "paddle/fluid/framework/details/build_strategy.h"
 #include "paddle/fluid/framework/details/execution_strategy.h"
 #include "paddle/fluid/framework/executor.h"
diff --git a/paddle/fluid/framework/scope.cc b/paddle/fluid/framework/scope.cc
index 0d261dd7ccc323abddd2c3ef13f1874661a8ca75..873f68e42e4fb2b839f64da001014dab8af790da 100644
--- a/paddle/fluid/framework/scope.cc
+++ b/paddle/fluid/framework/scope.cc
@@ -58,7 +58,10 @@ int64_t GetEagerDeletionThreshold() {
                                     (static_cast<int64_t>(1) << 30));
 }
 
-Scope::~Scope() { DropKids(); }
+Scope::~Scope() {
+  VLOG(5) << "~Scope()";
+  DropKids();
+}
 
 Scope& Scope::NewScope() const {
   SCOPE_LOCK_GUARD
diff --git a/paddle/fluid/framework/threadpool.cc b/paddle/fluid/framework/threadpool.cc
index fcec955360f1c681a62929e904d5736854a8ffad..7dc7430c55b8147a11441b9a928b4712822e3a98 100644
--- a/paddle/fluid/framework/threadpool.cc
+++ b/paddle/fluid/framework/threadpool.cc
@@ -48,9 +48,18 @@ void ThreadPool::Init() {
 
 ThreadPool::ThreadPool(int num_threads) : running_(true) {
   threads_.resize(num_threads);
-  for (auto& thread : threads_) {
+  for (int i = 0; i < num_threads; ++i) {
+    // for (auto& thread : threads_) {
     // TODO(Yancey1989): binding the thread on the specify CPU number
-    thread.reset(new std::thread(std::bind(&ThreadPool::TaskLoop, this)));
+    threads_[i].reset(
+        new std::thread(std::bind(&ThreadPool::TaskLoop, this, i)));
+    /**
+    sched_param sch;
+    int policy;
+    pthread_getschedparam(threads_[i]->native_handle(), &policy, &sch);
+    if (pthread_setschedparam(threads_[i]->native_handle(), SCHED_FIFO, &sch)) {
+      VLOG(1) << "Failed to setschedparam: " << errno;
+    }**/
   }
 }
 
@@ -68,7 +77,7 @@ ThreadPool::~ThreadPool() {
   }
 }
 
-void ThreadPool::TaskLoop() {
+void ThreadPool::TaskLoop(int i) {
   while (true) {
     Task task;
 
@@ -89,7 +98,6 @@ void ThreadPool::TaskLoop() {
       task = std::move(tasks_.front());
       tasks_.pop();
     }
-
     // run the task
     task();
   }
diff --git a/paddle/fluid/framework/threadpool.h b/paddle/fluid/framework/threadpool.h
index 7a51d18fbbf65f68725aa86a6a0ce4d15dff5673..bd8c3cdee8ccbc7e110c2be0310693e2595d272b 100644
--- a/paddle/fluid/framework/threadpool.h
+++ b/paddle/fluid/framework/threadpool.h
@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <pthread.h>
 #include <condition_variable>  // NOLINT
 #include <functional>
 #include <future>  // NOLINT
@@ -27,7 +28,6 @@ limitations under the License. */
 
 namespace paddle {
 namespace framework {
-
 struct ExceptionHandler {
   mutable std::future<std::unique_ptr<platform::EnforceNotMet>> future_;
   explicit ExceptionHandler(
@@ -99,7 +99,7 @@ class ThreadPool {
 
   // The constructor starts threads to run TaskLoop, which retrieves
   // and runs tasks from the queue.
-  void TaskLoop();
+  void TaskLoop(int i);
 
   // Init is called by GetInstance.
   static void Init();
diff --git a/paddle/fluid/framework/threadpool_test.cc b/paddle/fluid/framework/threadpool_test.cc
index 884d61e23428a0ad758946295ca9c470767e93ef..1257a76e3e64fb7dfa0b1f1095b2e939bdae9b49 100644
--- a/paddle/fluid/framework/threadpool_test.cc
+++ b/paddle/fluid/framework/threadpool_test.cc
@@ -59,3 +59,47 @@ TEST(ThreadPool, ConcurrentRun) {
   }
   EXPECT_EQ(sum, ((n + 1) * n) / 2);
 }
+static int64_t GetTS() {
+  struct timeval tp;
+  gettimeofday(&tp, NULL);
+  return tp.tv_sec * 1000000 + tp.tv_usec;
+}
+
+void multi_call(std::function<void()> call) {
+  for (int i = 0; i < 500; ++i) {
+    call();
+  }
+}
+
+TEST(ThreadPool, PERFORMANCE) {
+  auto sum = [] {
+    int a = 0;
+    for (int i = 0; i < 1000; ++i) {
+      a += i;
+    }
+  };
+  // framework::ThreadPool *pool = new framework::ThreadPool(2);
+  int64_t start = GetTS();
+  for (int i = 0; i < 1000; ++i) {
+    // int64_t s = GetTS();
+    framework::Async(std::move(sum));
+    // pool->Run(std::move(sum));
+    // VLOG(5) << "push to pool spent : " << GetTS() - s << " (us).";
+  }
+  VLOG(5) << "pool spent: " << GetTS() - start << " (us).";
+  start = GetTS();
+  for (int i = 0; i < 1000; ++i) {
+    sum();
+  }
+  VLOG(5) << "sequence call spent: " << GetTS() - start << " (us).";
+  std::vector<std::thread> threads;
+  start = GetTS();
+  for (int i = 0; i < 2; ++i) {
+    std::thread t(multi_call, std::ref(sum));
+    threads.push_back(std::move(t));
+  }
+  for (auto& thread : threads) {
+    thread.join();
+  }
+  VLOG(5) << "two threads spent: " << GetTS() - start << " (us).";
+}
diff --git a/paddle/fluid/operators/reader/blocking_queue.h b/paddle/fluid/operators/reader/blocking_queue.h
index 51b980acb5a08d431d96a3a92479dec09119c27e..10de11bfa5c16f71c131a3352b6a3d63915824e3 100644
--- a/paddle/fluid/operators/reader/blocking_queue.h
+++ b/paddle/fluid/operators/reader/blocking_queue.h
@@ -67,9 +67,12 @@ class BlockingQueue {
   }
 
   bool Receive(T* elem) {
+    VLOG(1) << "blocking queue::Receive ...";
     std::unique_lock<std::mutex> lock(mutex_);
     receive_cv_.wait(lock, [&] { return !queue_.empty() || closed_; });
+    VLOG(1) << "queue_.empty()=" << queue_.empty();
     if (!queue_.empty()) {
+      if (elem == nullptr) VLOG(1) << "elem is nullptr";
       PADDLE_ENFORCE_NOT_NULL(elem);
       *elem = queue_.front();
       if (LIKELY(!speed_test_mode_)) {
diff --git a/paddle/fluid/operators/reader/buffered_reader.cc b/paddle/fluid/operators/reader/buffered_reader.cc
index 26ff221dfa0768bd2bcc9e6485a32485f0212ac6..2d66000f1f88169a6b03217b80ea4770f27d5b55 100644
--- a/paddle/fluid/operators/reader/buffered_reader.cc
+++ b/paddle/fluid/operators/reader/buffered_reader.cc
@@ -58,7 +58,9 @@ void BufferedReader::ReadAsync(size_t i) {
       TensorVec &gpu = gpu_buffer_[i];
       gpu.resize(cpu.size());
       for (size_t i = 0; i < cpu.size(); ++i) {
+        VLOG(1) << "launch tensor copy from cpu to cpu, idx: " << i;
         framework::TensorCopySync(cpu[i], place_, &gpu[i]);
+        VLOG(1) << "done " << i;
         gpu[i].set_lod(cpu[i].lod());
       }
     }
@@ -80,11 +82,13 @@ void BufferedReader::StartImpl() {
 }
 
 void BufferedReader::ReadNextImpl(std::vector<framework::LoDTensor> *out) {
+  VLOG(1) << "ReadNextImpl start on place: " << place_;
   if (position_.empty()) {
     out->clear();
     return;
   }
   size_t i = position_.front().get();
+  VLOG(1) << "position front: " << i;
   position_.pop();
 
   if (i == -1UL) {
@@ -101,6 +105,7 @@ void BufferedReader::ReadNextImpl(std::vector<framework::LoDTensor> *out) {
     ReadAsync(prev_pos_);
   }
   prev_pos_ = i;
+  VLOG(1) << "success ReadNextImpl";
 }
 
 }  // namespace reader
diff --git a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc
index ed719f91d0980480aa62a5cd3c1f819e6c0e7475..924c92e0bfdf764d7d02dafb2fe724611a1a3ed0 100644
--- a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc
+++ b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc
@@ -25,9 +25,15 @@ class CreateDoubleBufferReaderOp : public framework::OperatorBase {
  private:
   void RunImpl(const framework::Scope& scope,
                const platform::Place& dev_place) const override {
-    auto* out = scope.FindVar(Output("Out"))
-                    ->template GetMutable<framework::ReaderHolder>();
+    VLOG(1) << "find var in scope: " << &scope;
+    auto* out_var = scope.FindVar(Output("Out"));
+    VLOG(1) << "var " << Output("Out") << " -> " << out_var;
+    auto* out = out_var->GetMutable<framework::ReaderHolder>();
+
+    // auto* out = scope.Var(Output("Out"))
+    //                ->template GetMutable<framework::ReaderHolder>();
     if (out->Get() != nullptr) {
+      VLOG(1) << Output("Out") << " is not nullptr.";
       return;
     }
     const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader"))
@@ -46,9 +52,11 @@ class CreateDoubleBufferReaderOp : public framework::OperatorBase {
       sin >> num;
       place = platform::CUDAPlace(static_cast<int>(num));
     }
-
+    VLOG(1) << "create buffered reader on " << place;
     out->Reset(framework::MakeDecoratedReader<BufferedReader>(underlying_reader,
                                                               place, 2));
+    VLOG(1) << "Reset Buffered Reader in var: "
+            << scope.FindVar(Input("UnderlyingReader"));
   }
 };
 
diff --git a/paddle/fluid/operators/reader/create_py_reader_op.cc b/paddle/fluid/operators/reader/create_py_reader_op.cc
index 901a92ab5b5c74b071be8b57a7653d90e2a4fb29..093b0e56b3d975ea02d3248b6a4d4c734231856b 100644
--- a/paddle/fluid/operators/reader/create_py_reader_op.cc
+++ b/paddle/fluid/operators/reader/create_py_reader_op.cc
@@ -28,8 +28,10 @@ class PyReader : public framework::FileReader {
   }
 
   void ReadNext(std::vector<framework::LoDTensor>* out) override {
+    VLOG(1) << "come in PyReader::ReadNext function, out: " << out;
     bool success;
     *out = queue_->Pop(&success);
+    VLOG(1) << "call PyReader::ReadNext " << success;
     if (!success) out->clear();
   }
 
diff --git a/paddle/fluid/operators/reader/open_files_op.cc b/paddle/fluid/operators/reader/open_files_op.cc
index 38223e069975a08791d58d6ae10e2112b79a61fe..ae37a187259e10388d83d5cf795f9474ee0b7962 100644
--- a/paddle/fluid/operators/reader/open_files_op.cc
+++ b/paddle/fluid/operators/reader/open_files_op.cc
@@ -115,10 +115,12 @@ class PreemptiveReaderContainer : public IReaderContainer {
   }
 
   void ReadNext(std::vector<framework::LoDTensor>* out) override {
+    VLOG(1) << "flag";
     if (!pending_.empty()) {
       auto future_it = complete_queue_.Pop();
       FutureItem item = future_it->get();
       if (item.exception_) {
+        VLOG(1) << "item has exception!!!";
         for (auto it = futures_.begin(); it != futures_.end(); ++it) {
           if (it != future_it) {
             it->wait();  // Wait all other threads complete.
diff --git a/paddle/fluid/platform/nccl_helper.h b/paddle/fluid/platform/nccl_helper.h
index 7c539d25f6dd02fc09aa1234d7bf0164b54a610f..53de53f43df2d0413933cf744d368f3d51729021 100644
--- a/paddle/fluid/platform/nccl_helper.h
+++ b/paddle/fluid/platform/nccl_helper.h
@@ -82,12 +82,15 @@ struct NCCLContext {
 struct NCCLContextMap {
   std::unordered_map<int, NCCLContext> contexts_;
   std::vector<int> order_;
+  bool need_group_call_;
 
   explicit NCCLContextMap(const std::vector<platform::Place> &places,
                           ncclUniqueId *nccl_id = nullptr,
-                          size_t num_trainers = 1, size_t trainer_id = 0) {
+                          size_t num_trainers = 1, size_t trainer_id = 0,
+                          bool need_group_call = true) {
     PADDLE_ENFORCE(!places.empty());
     order_.reserve(places.size());
+    need_group_call_ = need_group_call;
     for (auto &p : places) {
       int dev_id = boost::get<CUDAPlace>(p).device;
       order_.emplace_back(dev_id);
@@ -102,7 +105,7 @@ struct NCCLContextMap {
     }
     std::unique_ptr<ncclComm_t[]> comms(new ncclComm_t[order_.size()]);
     // if num_trainers == 1, should create a new nccl id for local comms.
-    if (num_trainers == 1) {
+    if (num_trainers == 1 && nccl_id != nullptr) {
       std::lock_guard<std::mutex> guard(NCCLGroupGuard::NCCLMutex());
       PADDLE_ENFORCE(platform::dynload::ncclCommInitAll(
           comms.get(), static_cast<int>(order_.size()), order_.data()));
diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc
index 998242fb4a09138db24aa75759f4990ffdc4d4e2..040a68f67264d0068fc484b6629828bd2bdd222a 100644
--- a/paddle/fluid/platform/profiler.cc
+++ b/paddle/fluid/platform/profiler.cc
@@ -12,9 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/platform/profiler.h"
-#include "paddle/fluid/platform/port.h"
-
 #include <algorithm>
 #include <iomanip>
 #include <limits>
@@ -25,9 +22,12 @@ limitations under the License. */
 #ifdef PADDLE_WITH_CUDA
 #include <cuda.h>
 #endif  // PADDLE_WITH_CUDA
+
 #include "glog/logging.h"
 #include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/platform/device_tracer.h"
+#include "paddle/fluid/platform/port.h"
+#include "paddle/fluid/platform/profiler.h"
 #include "paddle/fluid/string/printf.h"
 
 DEFINE_bool(enable_rpc_profiler, false, "Enable rpc profiler or not.");
@@ -173,8 +173,9 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx) {
 
 RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx)
     : is_enabled_(false), start_ns_(PosixInNsec()) {
-  std::lock_guard<std::mutex> l(profiler_mu);
   if (g_state == ProfilerState::kDisabled) return;
+  std::lock_guard<std::mutex> l(profiler_mu);
+
   is_enabled_ = true;
   dev_ctx_ = dev_ctx;
   name_ = name;
@@ -184,8 +185,9 @@ RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx)
 }
 
 RecordEvent::~RecordEvent() {
-  std::lock_guard<std::mutex> l(profiler_mu);
   if (g_state == ProfilerState::kDisabled || !is_enabled_) return;
+  VLOG(5) << "call ~RecordEvent";
+  std::lock_guard<std::mutex> l(profiler_mu);
   DeviceTracer* tracer = GetDeviceTracer();
   if (tracer) {
     tracer->AddCPURecords(CurAnnotation(), start_ns_, PosixInNsec(),
diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index fc7991d2974c9262e6225de1537025944c1068c1..c313ed2a8b95303cd4dc5fa38fc82ae0586a2f53 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -720,6 +720,11 @@ All parameter, weight, gradient are variables in Paddle.
 
         )DOC");
 
+  py::enum_<ExecutionStrategy::ExecutorType>(exec_strategy, "ExecutorType")
+      .value("Default", ExecutionStrategy::ExecutorType::kDefault)
+      .value("Experimental", ExecutionStrategy::ExecutorType::kExperimental)
+      .value("ParallelGraph", ExecutionStrategy::ExecutorType::kParallelGraph);
+
   exec_strategy.def(py::init())
       .def_property(
           "num_threads",
@@ -777,17 +782,14 @@ All parameter, weight, gradient are variables in Paddle.
                     [](const ExecutionStrategy &self) { return self.dry_run_; },
                     [](ExecutionStrategy &self, bool dry_run) {
                       self.dry_run_ = dry_run;
-                    });
-
-  exec_strategy.def_property(
-      "use_experimental_executor",
-      [](const ExecutionStrategy &self) {
-        return self.type_ == ExecutionStrategy::kExperimental;
-      },
-      [](ExecutionStrategy &self, bool experimental) {
-        self.type_ = experimental ? ExecutionStrategy::kExperimental
-                                  : ExecutionStrategy::kDefault;
-      });
+                    })
+      .def_property(
+          "executor_type",
+          [](const ExecutionStrategy &self) { return self.type_; },
+          [](ExecutionStrategy &self, ExecutionStrategy::ExecutorType type) {
+            self.type_ = type;
+          },
+          R"DOC()DOC");
 
   py::class_<BuildStrategy> build_strategy(pe, "BuildStrategy", R"DOC(
     BuildStrategy allows the user to more preciously control how to