diff --git a/python/examples/bert/benchmark.py b/python/examples/bert/benchmark.py
index af75b718b78b2bc130c2411d05d190fc0d298006..f1533d9710d3149a37818d3f1bc146fad6ce6537 100644
--- a/python/examples/bert/benchmark.py
+++ b/python/examples/bert/benchmark.py
@@ -21,11 +21,7 @@ import sys
 import time
 from paddle_serving_client import Client
 from paddle_serving_client.utils import MultiThreadRunner
-from paddle_serving_client.utils import benchmark_args
-from batching import pad_batch_data
-import tokenization
-import requests
-import json
+from paddle_serving_client.utils import benchmark_args, show_latency
 from paddle_serving_app.reader import ChineseBertReader
 
 args = benchmark_args()
@@ -36,42 +32,75 @@ def single_func(idx, resource):
     dataset = []
     for line in fin:
         dataset.append(line.strip())
+
+    profile_flags = False
+    latency_flags = False
+    if os.getenv("FLAGS_profile_client"):
+        profile_flags = True
+    if os.getenv("FLAGS_serving_latency"):
+        latency_flags = True
+        latency_list = []
+
     if args.request == "rpc":
-        reader = ChineseBertReader(vocab_file="vocab.txt", max_seq_len=20)
+        reader = ChineseBertReader({"max_seq_len": 128})
         fetch = ["pooled_output"]
         client = Client()
         client.load_client_config(args.model)
         client.connect([resource["endpoint"][idx % len(resource["endpoint"])]])
-
         start = time.time()
-        for i in range(1000):
-            if args.batch_size == 1:
-                feed_dict = reader.process(dataset[i])
-                result = client.predict(feed=feed_dict, fetch=fetch)
+        for i in range(turns):
+            if args.batch_size >= 1:
+                l_start = time.time()
+                feed_batch = []
+                b_start = time.time()
+                for bi in range(args.batch_size):
+                    feed_batch.append(reader.process(dataset[bi]))
+                b_end = time.time()
+
+                if profile_flags:
+                    sys.stderr.write(
+                        "PROFILE\tpid:{}\tbert_pre_0:{} bert_pre_1:{}\n".format(
+                            os.getpid(),
+                            int(round(b_start * 1000000)),
+                            int(round(b_end * 1000000))))
+                result = client.predict(feed=feed_batch, fetch=fetch)
+
+                l_end = time.time()
+                if latency_flags:
+                    latency_list.append(l_end * 1000 - l_start * 1000)
             else:
                 print("unsupport batch size {}".format(args.batch_size))
 
     elif args.request == "http":
-        start = time.time()
-        header = {"Content-Type": "application/json"}
-        for i in range(1000):
-            dict_data = {"words": dataset[i], "fetch": ["pooled_output"]}
-            r = requests.post(
-                'http://{}/bert/prediction'.format(resource["endpoint"][
-                    idx % len(resource["endpoint"])]),
-                data=json.dumps(dict_data),
-                headers=header)
+        raise ("not implemented")
     end = time.time()
-    return [[end - start]]
+    if latency_flags:
+        return [[end - start], latency_list]
+    else:
+        return [[end - start]]
 
 
 if __name__ == '__main__':
     multi_thread_runner = MultiThreadRunner()
-    endpoint_list = ["127.0.0.1:9292"]
-    result = multi_thread_runner.run(single_func, args.thread,
-                                     {"endpoint": endpoint_list})
+    endpoint_list = [
+        "127.0.0.1:9292", "127.0.0.1:9293", "127.0.0.1:9294", "127.0.0.1:9295"
+    ]
+    turns = 10
+    start = time.time()
+    result = multi_thread_runner.run(
+        single_func, args.thread, {"endpoint": endpoint_list,
+                                   "turns": turns})
+    end = time.time()
+    total_cost = end - start
+
     avg_cost = 0
     for i in range(args.thread):
         avg_cost += result[0][i]
     avg_cost = avg_cost / args.thread
-    print("average total cost {} s.".format(avg_cost))
+
+    print("total cost :{} s".format(total_cost))
+    print("each thread cost :{} s. ".format(avg_cost))
+    print("qps :{} samples/s".format(args.batch_size * args.thread * turns /
+                                     total_cost))
+    if os.getenv("FLAGS_serving_latency"):
+        show_latency(result[1])
diff --git a/python/examples/bert/benchmark.sh b/python/examples/bert/benchmark.sh
index 7f9e2325f3b8f7db288d2b7d82d0d412e05417cb..7ee5f32e9e5d89a836f8962a256bcdf7bf0b62e2 100644
--- a/python/examples/bert/benchmark.sh
+++ b/python/examples/bert/benchmark.sh
@@ -1,9 +1,30 @@
 rm profile_log
-for thread_num in 1 2 4 8 16
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+export FLAGS_profile_server=1
+export FLAGS_profile_client=1
+export FLAGS_serving_latency=1
+python3 -m paddle_serving_server_gpu.serve --model $1 --port 9292 --thread 4 --gpu_ids 0,1,2,3 --mem_optim False --ir_optim True 2> elog > stdlog &
+
+sleep 5
+
+#warm up
+python3 benchmark.py --thread 8 --batch_size 1 --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+
+for thread_num in 4 8 16
 do
-    $PYTHONROOT/bin/python benchmark.py --thread $thread_num --model serving_client_conf/serving_client_conf.prototxt --request rpc > profile 2>&1
-    echo "========================================"
-    echo "batch size : $batch_size" >> profile_log
-    $PYTHONROOT/bin/python ../util/show_profile.py profile $thread_num >> profile_log
-    tail -n 1 profile >> profile_log
+for batch_size in 1 4 16 64 256
+do
+    python3 benchmark.py --thread $thread_num --batch_size $batch_size --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+    echo "model name :" $1
+    echo "thread num :" $thread_num
+    echo "batch size :" $batch_size
+    echo "=================Done===================="
+    echo "model name :$1" >> profile_log_$1
+    echo "batch size :$batch_size" >> profile_log_$1
+    python3 ../util/show_profile.py profile $thread_num >> profile_log_$1
+    tail -n 8 profile >> profile_log_$1
+    echo "" >> profile_log_$1
+done
 done
+
+ps -ef|grep 'serving'|grep -v grep|cut -c 9-15 | xargs kill -9
diff --git a/python/examples/bert/benchmark_batch.py b/python/examples/bert/benchmark_batch.py
deleted file mode 100644
index 7cedb6aa451e0e4a128f0fedbfde1a896977f601..0000000000000000000000000000000000000000
--- a/python/examples/bert/benchmark_batch.py
+++ /dev/null
@@ -1,79 +0,0 @@
-# -*- coding: utf-8 -*-
-#
-# Copyright (c) 2020 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.
-# pylint: disable=doc-string-missing
-
-from __future__ import unicode_literals, absolute_import
-import os
-import sys
-import time
-from paddle_serving_client import Client
-from paddle_serving_client.utils import MultiThreadRunner
-from paddle_serving_client.utils import benchmark_args
-from batching import pad_batch_data
-import tokenization
-import requests
-import json
-from bert_reader import BertReader
-args = benchmark_args()
-
-
-def single_func(idx, resource):
-    fin = open("data-c.txt")
-    dataset = []
-    for line in fin:
-        dataset.append(line.strip())
-    profile_flags = False
-    if os.environ["FLAGS_profile_client"]:
-        profile_flags = True
-    if args.request == "rpc":
-        reader = BertReader(vocab_file="vocab.txt", max_seq_len=20)
-        fetch = ["pooled_output"]
-        client = Client()
-        client.load_client_config(args.model)
-        client.connect([resource["endpoint"][idx % len(resource["endpoint"])]])
-        start = time.time()
-        for i in range(1000):
-            if args.batch_size >= 1:
-                feed_batch = []
-                b_start = time.time()
-                for bi in range(args.batch_size):
-                    feed_batch.append(reader.process(dataset[bi]))
-                b_end = time.time()
-                if profile_flags:
-                    print("PROFILE\tpid:{}\tbert_pre_0:{} bert_pre_1:{}".format(
-                        os.getpid(),
-                        int(round(b_start * 1000000)),
-                        int(round(b_end * 1000000))))
-                result = client.predict(feed=feed_batch, fetch=fetch)
-            else:
-                print("unsupport batch size {}".format(args.batch_size))
-
-    elif args.request == "http":
-        raise ("no batch predict for http")
-    end = time.time()
-    return [[end - start]]
-
-
-if __name__ == '__main__':
-    multi_thread_runner = MultiThreadRunner()
-    endpoint_list = ["127.0.0.1:9292"]
-    result = multi_thread_runner.run(single_func, args.thread,
-                                     {"endpoint": endpoint_list})
-    avg_cost = 0
-    for i in range(args.thread):
-        avg_cost += result[0][i]
-    avg_cost = avg_cost / args.thread
-    print("average total cost {} s.".format(avg_cost))
diff --git a/python/examples/bert/benchmark_batch.sh b/python/examples/bert/benchmark_batch.sh
deleted file mode 100644
index 272923776d6640880175745920a8fad9e84972fd..0000000000000000000000000000000000000000
--- a/python/examples/bert/benchmark_batch.sh
+++ /dev/null
@@ -1,19 +0,0 @@
-rm profile_log
-export CUDA_VISIBLE_DEVICES=0,1,2,3
-python -m paddle_serving_server_gpu.serve --model bert_seq20_model/ --port 9295 --thread 4 --gpu_ids 0,1,2,3 2> elog > stdlog &
-
-sleep 5
-
-for thread_num in 1 2 4 8 16
-do
-for batch_size in 1 2 4 8 16 32 64 128 256 512
-do
-    $PYTHONROOT/bin/python benchmark_batch.py --thread $thread_num --batch_size $batch_size --model serving_client_conf/serving_client_conf.prototxt --request rpc > profile 2>&1
-    echo "========================================"
-    echo "thread num: ", $thread_num
-    echo "batch size: ", $batch_size
-    echo "batch size : $batch_size" >> profile_log
-    $PYTHONROOT/bin/python ../util/show_profile.py profile $thread_num >> profile_log
-    tail -n 1 profile >> profile_log
-done
-done
diff --git a/python/examples/imagenet/benchmark.py b/python/examples/imagenet/benchmark.py
index caa952f121fbd8725c2a6bfe36f0dd84b6a82707..ac7ba8c333d25fb23bfc7695105315bfaa4e76ee 100644
--- a/python/examples/imagenet/benchmark.py
+++ b/python/examples/imagenet/benchmark.py
@@ -93,7 +93,7 @@ def single_func(idx, resource):
 
 if __name__ == '__main__':
     multi_thread_runner = MultiThreadRunner()
-    endpoint_list = ["127.0.0.1:9696"]
+    endpoint_list = ["127.0.0.1:9393"]
     #endpoint_list = endpoint_list + endpoint_list + endpoint_list
     result = multi_thread_runner.run(single_func, args.thread,
                                      {"endpoint": endpoint_list})
diff --git a/python/examples/imagenet/benchmark.sh b/python/examples/imagenet/benchmark.sh
index 618a62c063c0bc4955baf8516bc5bc93e4832394..84885908fa89d050b3ca71386fe2a21533ce0809 100644
--- a/python/examples/imagenet/benchmark.sh
+++ b/python/examples/imagenet/benchmark.sh
@@ -1,12 +1,28 @@
 rm profile_log
-for thread_num in 1 2 4 8
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+export FLAGS_profile_server=1
+export FLAGS_profile_client=1
+python -m paddle_serving_server_gpu.serve --model $1 --port 9292 --thread 4 --gpu_ids 0,1,2,3 2> elog > stdlog &
+
+sleep 5
+
+#warm up
+$PYTHONROOT/bin/python benchmark.py --thread 8 --batch_size 1 --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+
+for thread_num in 4 8 16
 do
-for batch_size in 1 2 4 8 16 32 64 128
+for batch_size in 1 4 16 64 256
 do
-    $PYTHONROOT/bin/python benchmark.py --thread $thread_num --batch_size $batch_size --model ResNet50_vd_client_config/serving_client_conf.prototxt --request rpc > profile 2>&1
-    echo "========================================"
-    echo "batch size : $batch_size" >> profile_log
+    $PYTHONROOT/bin/python benchmark.py --thread $thread_num --batch_size $batch_size --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+    echo "model name :" $1
+    echo "thread num :" $thread_num
+    echo "batch size :" $batch_size
+    echo "=================Done===================="
+    echo "model name :$1" >> profile_log
+    echo "batch size :$batch_size" >> profile_log
     $PYTHONROOT/bin/python ../util/show_profile.py profile $thread_num >> profile_log
-    tail -n 1 profile >> profile_log
+    tail -n 8 profile >> profile_log
 done
 done
+
+ps -ef|grep 'serving'|grep -v grep|cut -c 9-15 | xargs kill -9
diff --git a/python/examples/util/show_profile.py b/python/examples/util/show_profile.py
index 9153d939338f0ee171af539b9f955d51802ad547..1581dda19bb0abefe6eb21592bda7fc97d8fb7cd 100644
--- a/python/examples/util/show_profile.py
+++ b/python/examples/util/show_profile.py
@@ -31,7 +31,7 @@ with open(profile_file) as f:
         if line[0] == "PROFILE":
             prase(line[2])
 
-print("thread num {}".format(thread_num))
+print("thread num :{}".format(thread_num))
 for name in time_dict:
-    print("{} cost {} s in each thread ".format(name, time_dict[name] / (
+    print("{} cost :{} s in each thread ".format(name, time_dict[name] / (
         1000000.0 * float(thread_num))))
diff --git a/python/paddle_serving_client/utils/__init__.py b/python/paddle_serving_client/utils/__init__.py
index 381da6bf9bade2bb0627f4c07851012360905de5..53f40726fbf21a0607b47bb29a20aa6ff50b6221 100644
--- a/python/paddle_serving_client/utils/__init__.py
+++ b/python/paddle_serving_client/utils/__init__.py
@@ -17,6 +17,7 @@ import sys
 import subprocess
 import argparse
 from multiprocessing import Pool
+import numpy as np
 
 
 def benchmark_args():
@@ -35,6 +36,17 @@ def benchmark_args():
     return parser.parse_args()
 
 
+def show_latency(latency_list):
+    latency_array = np.array(latency_list)
+    info = "latency:\n"
+    info += "mean :{} ms\n".format(np.mean(latency_array))
+    info += "median :{} ms\n".format(np.median(latency_array))
+    info += "80 percent :{} ms\n".format(np.percentile(latency_array, 80))
+    info += "90 percent :{} ms\n".format(np.percentile(latency_array, 90))
+    info += "99 percent :{} ms\n".format(np.percentile(latency_array, 99))
+    sys.stderr.write(info)
+
+
 class MultiThreadRunner(object):
     def __init__(self):
         pass