diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt
index 2c5364b72402befd2c34e5f542ce5c6b2add621d..058a5b5f460d2bd3c4c0248929dd0c87f7506930 100644
--- a/paddle/fluid/inference/CMakeLists.txt
+++ b/paddle/fluid/inference/CMakeLists.txt
@@ -4,6 +4,7 @@ endif()
 # analysis and tensorrt must be added before creating static library,
 # otherwise, there would be undefined reference to them in static library.
 add_subdirectory(analysis)
+add_subdirectory(utils)
 if (TENSORRT_FOUND)
   add_subdirectory(tensorrt)
 endif()
diff --git a/paddle/fluid/inference/utils/CMakeLists.txt b/paddle/fluid/inference/utils/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..2104e4ac7222258ee025bd5acd60b1db251df654
--- /dev/null
+++ b/paddle/fluid/inference/utils/CMakeLists.txt
@@ -0,0 +1,2 @@
+cc_library(benchmark SRCS benchmark.cc DEPS enforce)
+cc_test(test_benchmark SRCS benchmark_tester.cc DEPS benchmark)
diff --git a/paddle/fluid/inference/utils/benchmark.cc b/paddle/fluid/inference/utils/benchmark.cc
new file mode 100644
index 0000000000000000000000000000000000000000..021edc2de5e90023fcd1431dd2025450e7462bd9
--- /dev/null
+++ b/paddle/fluid/inference/utils/benchmark.cc
@@ -0,0 +1,49 @@
+// 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/inference/utils/benchmark.h"
+#include <sstream>
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace inference {
+
+std::string Benchmark::SerializeToString() const {
+  std::stringstream ss;
+  ss << "-----------------------------------------------------\n";
+  ss << "name\t";
+  ss << "batch_size\t";
+  ss << "num_threads\t";
+  ss << "latency\t";
+  ss << "qps";
+  ss << '\n';
+
+  ss << name_ << "\t";
+  ss << batch_size_ << "\t";
+  ss << num_threads_ << "\t";
+  ss << latency_ << "\t";
+  ss << 1000 / latency_;
+  ss << '\n';
+  return ss.str();
+}
+void Benchmark::PersistToFile(const std::string &path) const {
+  std::ofstream file(path, std::ios::app);
+  PADDLE_ENFORCE(file.is_open(), "Can not open %s to add benchmark", path);
+  file << SerializeToString();
+  file.flush();
+  file.close();
+}
+
+}  // namespace inference
+}  // namespace paddle
diff --git a/paddle/fluid/inference/utils/benchmark.h b/paddle/fluid/inference/utils/benchmark.h
new file mode 100644
index 0000000000000000000000000000000000000000..80e8f77adb4ff2cc81a2a3dd0c44e4e304800122
--- /dev/null
+++ b/paddle/fluid/inference/utils/benchmark.h
@@ -0,0 +1,52 @@
+// 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 <fstream>
+#include <iostream>
+
+namespace paddle {
+namespace inference {
+
+/*
+ * Helper class to calculate the performance.
+ */
+struct Benchmark {
+  int batch_size() const { return batch_size_; }
+  void SetBatchSize(int x) { batch_size_ = x; }
+
+  int num_threads() const { return num_threads_; }
+  void SetNumThreads(int x) { num_threads_ = x; }
+
+  bool use_gpu() const { return use_gpu_; }
+  void SetUseGpu() { use_gpu_ = true; }
+
+  int latency() const { return latency_; }
+  void SetLatency(int x) { latency_ = x; }
+
+  const std::string& name() const { return name_; }
+  void SetName(const std::string& name) { name_ = name; }
+
+  std::string SerializeToString() const;
+  void PersistToFile(const std::string& path) const;
+
+ private:
+  bool use_gpu_{false};
+  int batch_size_{0};
+  int latency_;
+  int num_threads_{1};
+  std::string name_;
+};
+
+}  // namespace inference
+}  // namespace paddle
diff --git a/paddle/fluid/inference/utils/benchmark_tester.cc b/paddle/fluid/inference/utils/benchmark_tester.cc
new file mode 100644
index 0000000000000000000000000000000000000000..eb255474082b27180a8b3176b5f880c0d38f6c3b
--- /dev/null
+++ b/paddle/fluid/inference/utils/benchmark_tester.cc
@@ -0,0 +1,39 @@
+// 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/inference/utils/benchmark.h"
+#include <glog/logging.h>
+#include <gtest/gtest.h>
+
+using namespace paddle::inference;
+TEST(Benchmark, basic) {
+  Benchmark benchmark;
+  benchmark.SetName("key0");
+  benchmark.SetBatchSize(10);
+  benchmark.SetUseGpu();
+  benchmark.SetLatency(220);
+  LOG(INFO) << "benchmark:\n" << benchmark.SerializeToString();
+}
+
+TEST(Benchmark, PersistToFile) {
+  Benchmark benchmark;
+  benchmark.SetName("key0");
+  benchmark.SetBatchSize(10);
+  benchmark.SetUseGpu();
+  benchmark.SetLatency(220);
+
+  benchmark.PersistToFile("1.log");
+  benchmark.PersistToFile("1.log");
+  benchmark.PersistToFile("1.log");
+}
\ No newline at end of file