diff --git a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc
index d4fbea9d95118666ababde811867e95c657c07de..297ee92fc3c84c2feec9cb85bd8671ce8ad94ed0 100644
--- a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc
+++ b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc
@@ -31,9 +31,10 @@ FastThreadedSSAGraphExecutor::FastThreadedSSAGraphExecutor(
       local_scopes_(local_scopes),
       places_(places),
       graph_(graph),
+      fetch_ctxs_(places),
       pool_(strategy.num_threads_),
-      prepare_pool_(1),  // add one more thread for generate op_deps
-      fetch_ctxs_(places) {
+      // add one more thread for generate op_deps
+      prepare_pool_(1) {
   for (auto &op : ir::FilterByNodeWrapper<OpHandleBase>(*graph_)) {
     int dep = static_cast<int>(op->NotReadyInputSize());
     op_deps_.emplace(op, dep);
diff --git a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h
index 970298950cc8089bc5861fcbf8dc2544934b181f..f6d5160e75cc3f48c5129dae05eec4ec82d83ae5 100644
--- a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h
+++ b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h
@@ -14,7 +14,9 @@
 
 #pragma once
 #include <ThreadPool.h>
+#include <memory>
 #include <string>
+#include <unordered_map>
 #include <vector>
 #include "paddle/fluid/framework/blocking_queue.h"
 #include "paddle/fluid/framework/details/exception_holder.h"
@@ -37,6 +39,8 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
   const ir::Graph &Graph() const override;
 
  private:
+  // Note(zcd): the ThreadPool should be placed last so that ThreadPool should
+  // be destroyed first.
   ExecutionStrategy strategy_;
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
@@ -45,21 +49,22 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
   std::unordered_map<OpHandleBase *, int> op_deps_;
   std::vector<OpHandleBase *> bootstrap_ops_;
 
-  ::ThreadPool pool_;
-  ::ThreadPool prepare_pool_;
   platform::DeviceContextPool fetch_ctxs_;
   std::atomic<int> remaining_;
 
+  std::future<
+      std::unique_ptr<std::unordered_map<OpHandleBase *, std::atomic<int>>>>
+      atomic_op_deps_;
+  ExceptionHolder exception_;
+
+  ::ThreadPool pool_;
+  ::ThreadPool prepare_pool_;
+
   void RunOpAsync(std::unordered_map<OpHandleBase *, std::atomic<int>> *op_deps,
                   OpHandleBase *op,
                   const std::shared_ptr<BlockingQueue<size_t>> &complete_q);
 
   void PrepareAtomicOpDeps();
-
-  std::future<
-      std::unique_ptr<std::unordered_map<OpHandleBase *, std::atomic<int>>>>
-      atomic_op_deps_;
-  ExceptionHolder exception_;
 };
 }  // namespace details
 }  // namespace framework
diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
index c4254bbadfa17682f437f46f02adc9c884d24304..c00932a7bdb170e63b5fd4d43ccb2072f1a0a9c9 100644
--- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
+++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
@@ -24,13 +24,13 @@ ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor(
     const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places, ir::Graph *graph)
     : graph_(graph),
-      pool_(strategy.num_threads_ >= 2 ? new ::ThreadPool(strategy.num_threads_)
-                                       : nullptr),
-      prepare_pool_(1),
       local_scopes_(local_scopes),
       places_(places),
       fetch_ctxs_(places),
-      strategy_(strategy) {
+      strategy_(strategy),
+      prepare_pool_(1),
+      pool_(strategy.num_threads_ >= 2 ? new ::ThreadPool(strategy.num_threads_)
+                                       : nullptr) {
   PrepareOpDeps();
   CopyOpDeps();
 }
diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h
index b9bccba8fa2fa13d99a9a39a5135106101daa903..1fa5196970512ccc4a3dee698f477711be1e7101 100644
--- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h
+++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h
@@ -63,13 +63,20 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
              details::OpHandleBase *op);
 
  private:
+  // Note(zcd): the ThreadPool should be placed last so that ThreadPool should
+  // be destroyed first.
   ir::Graph *graph_;
-  std::unique_ptr<::ThreadPool> pool_;
-  ::ThreadPool prepare_pool_;
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
   platform::DeviceContextPool fetch_ctxs_;
   ExceptionHolder exception_holder_;
+  std::unique_ptr<OpDependentData> op_deps_;
+  std::future<std::unique_ptr<OpDependentData>> op_deps_futures_;
+  ExecutionStrategy strategy_;
+  // use std::list because clear(), push_back, and for_each are O(1)
+  std::list<std::future<void>> run_op_futures_;
+  ::ThreadPool prepare_pool_;
+  std::unique_ptr<::ThreadPool> pool_;
 
   void InsertPendingOp(std::unordered_map<OpHandleBase *, size_t> *pending_ops,
                        OpHandleBase *op_instance) const;
@@ -88,14 +95,6 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
 
   void PrepareOpDeps();
   void CopyOpDeps();
-
- private:
-  std::future<std::unique_ptr<OpDependentData>> op_deps_futures_;
-
-  ExecutionStrategy strategy_;
-  std::unique_ptr<OpDependentData> op_deps_;
-  // use std::list because clear(), push_back, and for_each are O(1)
-  std::list<std::future<void>> run_op_futures_;
 };
 
 }  // namespace details