diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
index 9f930065324f13f5aa79c214e820fb6fc2f3a166..14fcde2fe3b1c3acfc0994e9cd37a784c57826d7 100644
--- a/paddle/fluid/framework/operator.cc
+++ b/paddle/fluid/framework/operator.cc
@@ -149,9 +149,17 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
     platform::SetDeviceId(dev_id);
 #endif
   }
-  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-  platform::RecordEvent record_event(Type(), pool.Get(place));
-  RunImpl(scope, place);
+
+  // The profile has a process-wide mutex, results in serious performance issue
+  // in concurrency scenerio. Here use an `if` to fix this issue.
+  // Please not remove the `if`, ask @Superjomn if there are any concern.
+  if (platform::IsProfileEnabled()) {
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    platform::RecordEvent record_event(Type(), pool.Get(place));
+    RunImpl(scope, place);
+  } else {
+    RunImpl(scope, place);
+  }
   VLOG(3) << place << " " << DebugStringEx(&scope);
 }
 
diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
index f9135ff9d7be16b7dfb08c044188c1d7a567da4e..3095dee0f0106b2408663cd32bb4fb310111eda4 100644
--- a/paddle/fluid/inference/api/analysis_predictor.cc
+++ b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -340,6 +340,19 @@ bool AnalysisPredictor::LoadProgramDesc() {
   }
   return true;
 }
+
+AnalysisPredictor::~AnalysisPredictor() {
+#if !defined(_WIN32)
+  if (FLAGS_profile) {
+    platform::DisableProfiler(platform::EventSortingKey::kTotal,
+                              "./profile.log");
+  }
+#endif
+  if (sub_scope_) {
+    scope_->DeleteScope(sub_scope_);
+  }
+}
+
 std::unique_ptr<PaddlePredictor> AnalysisPredictor::Clone() {
   auto *x = new AnalysisPredictor(config_);
   x->Init(scope_, inference_program_);
diff --git a/paddle/fluid/inference/api/analysis_predictor.h b/paddle/fluid/inference/api/analysis_predictor.h
index 0d01d7ac2b29ea6364b07af9bb3bdeb5ced6bd00..5a9f4d36959d4ee7ca16dec769d9d1283b8787cb 100644
--- a/paddle/fluid/inference/api/analysis_predictor.h
+++ b/paddle/fluid/inference/api/analysis_predictor.h
@@ -72,6 +72,7 @@ class AnalysisPredictor : public PaddlePredictor {
   template <typename T>
   void GetFetchOne(const framework::LoDTensor &fetchs,
                    PaddleTensor *output_data);
+  ~AnalysisPredictor();
 
  private:
   contrib::AnalysisConfig config_;