diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc
index 2898c5ffd973b0d76bee4b9e3a24a8fce912fad2..875b5d8ba7acc826e6d657f96f1b26acb06141f6 100644
--- a/paddle/fluid/framework/parallel_executor.cc
+++ b/paddle/fluid/framework/parallel_executor.cc
@@ -702,7 +702,7 @@ void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
   while (!pending_ops.empty()) {
     VarHandleBase *ready_var = nullptr;
     for (auto &pair : pending_vars) {
-      if (pair.second) {
+      if (pair.second.load(std::memory_order_acquire)) {
         ready_var = pair.first;
       }
     }
@@ -714,7 +714,6 @@ void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
         throw * member_->exception_;
       }
 
-      std::this_thread::yield();
       continue;
     }
 
@@ -753,7 +752,7 @@ void ParallelExecutor::RunOp(
       VLOG(10) << op->DebugString();
       op->Run();
       for (auto *ready : ready_buffer) {
-        *ready = true;
+        ready->store(true, std::memory_order_release);
       }
     } catch (platform::EnforceNotMet ex) {
       member_->exception_.reset(new platform::EnforceNotMet(ex));
@@ -762,8 +761,7 @@ void ParallelExecutor::RunOp(
     }
   };
 
-  op_run();
-  //  member_->pool_.enqueue(op_run);
+  member_->pool_.enqueue(op_run);
 }
 }  // namespace framework
 }  // namespace paddle