diff --git a/python/examples/pipeline/PaddleClas/HRNet_W18_C/benchmark.sh b/python/examples/pipeline/PaddleClas/HRNet_W18_C/benchmark.sh
index ee20eccf64c1e08dff16c7427ae5107525281dee..9abbd0b362e6b85dc17ac29dd60eac8b88bbbb18 100644
--- a/python/examples/pipeline/PaddleClas/HRNet_W18_C/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/HRNet_W18_C/benchmark.sh
@@ -1,6 +1,6 @@
 export FLAGS_profile_pipeline=1
 alias python3="python3.6"
-modelname="clas-DarkNet53"
+modelname="clas-HRNet_W18_C"
 
 # HTTP
 #ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
diff --git a/python/examples/pipeline/PaddleClas/MobileNetV1/benchmark.sh b/python/examples/pipeline/PaddleClas/MobileNetV1/benchmark.sh
index ee20eccf64c1e08dff16c7427ae5107525281dee..d10142449db8b2ad3ea39e6e7531158b833b438a 100644
--- a/python/examples/pipeline/PaddleClas/MobileNetV1/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/MobileNetV1/benchmark.sh
@@ -1,6 +1,6 @@
 export FLAGS_profile_pipeline=1
 alias python3="python3.6"
-modelname="clas-DarkNet53"
+modelname="clas-MobileNetV1"
 
 # HTTP
 #ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
diff --git a/python/examples/pipeline/PaddleClas/MobileNetV3_large_x1_0/benchmark.sh b/python/examples/pipeline/PaddleClas/MobileNetV3_large_x1_0/benchmark.sh
index ee20eccf64c1e08dff16c7427ae5107525281dee..6a431f10ac7bdd05266e59318caafae4c223f427 100644
--- a/python/examples/pipeline/PaddleClas/MobileNetV3_large_x1_0/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/MobileNetV3_large_x1_0/benchmark.sh
@@ -1,6 +1,6 @@
 export FLAGS_profile_pipeline=1
 alias python3="python3.6"
-modelname="clas-DarkNet53"
+modelname="clas-MobileNetV3_large_x1_0"
 
 # HTTP
 #ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
diff --git a/python/examples/pipeline/PaddleClas/ResNeXt101_vd_64x4d/benchmark.sh b/python/examples/pipeline/PaddleClas/ResNeXt101_vd_64x4d/benchmark.sh
index ee20eccf64c1e08dff16c7427ae5107525281dee..2b9ec90e65a2c58e1b4b6b5c0237c6007e49c71e 100644
--- a/python/examples/pipeline/PaddleClas/ResNeXt101_vd_64x4d/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/ResNeXt101_vd_64x4d/benchmark.sh
@@ -1,6 +1,6 @@
 export FLAGS_profile_pipeline=1
 alias python3="python3.6"
-modelname="clas-DarkNet53"
+modelname="clas-ResNeXt101_vd_64x4d"
 
 # HTTP
 #ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
diff --git a/python/examples/pipeline/PaddleClas/ResNet50_vd/benchmark.sh b/python/examples/pipeline/PaddleClas/ResNet50_vd/benchmark.sh
index ee20eccf64c1e08dff16c7427ae5107525281dee..bc1c65244658912f6c700e3c342cec56813de4d3 100644
--- a/python/examples/pipeline/PaddleClas/ResNet50_vd/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/ResNet50_vd/benchmark.sh
@@ -1,6 +1,6 @@
 export FLAGS_profile_pipeline=1
 alias python3="python3.6"
-modelname="clas-DarkNet53"
+modelname="clas-ResNet50_vd"
 
 # HTTP
 #ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
diff --git a/python/examples/pipeline/PaddleClas/ResNet50_vd_FPGM/benchmark.sh b/python/examples/pipeline/PaddleClas/ResNet50_vd_FPGM/benchmark.sh
index ee20eccf64c1e08dff16c7427ae5107525281dee..17593214deccfa2f1af070ebd71c82775273019c 100644
--- a/python/examples/pipeline/PaddleClas/ResNet50_vd_FPGM/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/ResNet50_vd_FPGM/benchmark.sh
@@ -1,6 +1,6 @@
 export FLAGS_profile_pipeline=1
 alias python3="python3.6"
-modelname="clas-DarkNet53"
+modelname="clas-ResNet50_vd_FPGM"
 
 # HTTP
 #ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
diff --git a/python/examples/pipeline/PaddleClas/ResNet50_vd_KL/benchmark.sh b/python/examples/pipeline/PaddleClas/ResNet50_vd_KL/benchmark.sh
index ee20eccf64c1e08dff16c7427ae5107525281dee..248f9506e346aa15fd347f91f811e95657fdb66b 100644
--- a/python/examples/pipeline/PaddleClas/ResNet50_vd_KL/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/ResNet50_vd_KL/benchmark.sh
@@ -1,6 +1,6 @@
 export FLAGS_profile_pipeline=1
 alias python3="python3.6"
-modelname="clas-DarkNet53"
+modelname="clas-ResNet50_vd_KL"
 
 # HTTP
 #ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
diff --git a/python/examples/pipeline/PaddleClas/ResNet50_vd_PACT/benchmark.sh b/python/examples/pipeline/PaddleClas/ResNet50_vd_PACT/benchmark.sh
index ee20eccf64c1e08dff16c7427ae5107525281dee..06cbe71821afe257202c5088880d467053d6d762 100644
--- a/python/examples/pipeline/PaddleClas/ResNet50_vd_PACT/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/ResNet50_vd_PACT/benchmark.sh
@@ -1,6 +1,6 @@
 export FLAGS_profile_pipeline=1
 alias python3="python3.6"
-modelname="clas-DarkNet53"
+modelname="clas-ResNet50_vd_PACT"
 
 # HTTP
 #ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
diff --git a/python/examples/pipeline/PaddleClas/ShuffleNetV2_x1_0/benchmark.sh b/python/examples/pipeline/PaddleClas/ShuffleNetV2_x1_0/benchmark.sh
index ee20eccf64c1e08dff16c7427ae5107525281dee..08523c4a5ef929ffb5ce9c30f3c1059799a4f9f0 100644
--- a/python/examples/pipeline/PaddleClas/ShuffleNetV2_x1_0/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/ShuffleNetV2_x1_0/benchmark.sh
@@ -1,6 +1,6 @@
 export FLAGS_profile_pipeline=1
 alias python3="python3.6"
-modelname="clas-DarkNet53"
+modelname="clas-ShuffleNetV2_x1_0"
 
 # HTTP
 #ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
diff --git a/python/examples/pipeline/PaddleDetection/faster_rcnn/benchmark.py b/python/examples/pipeline/PaddleDetection/faster_rcnn/benchmark.py
index e2bef4f91e98ecd8459fc0ff431a9b784d3c639a..f0a55614c1390b1d4f73bd015b1ce21b85e4ba55 100644
--- a/python/examples/pipeline/PaddleDetection/faster_rcnn/benchmark.py
+++ b/python/examples/pipeline/PaddleDetection/faster_rcnn/benchmark.py
@@ -46,23 +46,43 @@ def run_http(idx, batch_size):
     with open(os.path.join(".", "000000570688.jpg"), 'rb') as file:
         image_data1 = file.read()
     image = cv2_to_base64(image_data1)
-
+    latency_list = []
     start = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         data = {"key": [], "value": []}
         for j in range(batch_size):
             data["key"].append("image_" + str(j))
             data["value"].append(image)
         r = requests.post(url=url, data=json.dumps(data))
+        l_end = time.time()
+        total_num += 1
         end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
         if end - start > 70:
-            print("70s end")
+            #print("70s end")
             break
-    return [[end - start]]
+    return [[end - start], latency_list, [total_num]]
 
 def multithread_http(thread, batch_size):
     multi_thread_runner = MultiThreadRunner()
-    result = multi_thread_runner.run(run_http , thread, batch_size)
+    start = time.time()
+    result = multi_thread_runner.run(run_http, thread, batch_size)
+    end = time.time()
+    total_cost = end - start
+    avg_cost = 0
+    total_number = 0
+    for i in range(thread):
+        avg_cost += result[0][i]
+        total_number += result[2][i]
+    avg_cost = avg_cost / thread
+    print("Total cost: {}s".format(total_cost))
+    print("Each thread cost: {}s. ".format(avg_cost))
+    print("Total count: {}. ".format(total_number))
+    print("AVG QPS: {} samples/s".format(batch_size * total_number /
+                                         total_cost))
+    show_latency(result[1])
 
 def run_rpc(thread, batch_size):
     pass
diff --git a/python/examples/pipeline/PaddleDetection/faster_rcnn/benchmark.sh b/python/examples/pipeline/PaddleDetection/faster_rcnn/benchmark.sh
index f67caaa7761929a48579b5d5ed56fa93bb3290ae..5247891f64f7a669a20e2ad19fa1f4cb94e1fb17 100644
--- a/python/examples/pipeline/PaddleDetection/faster_rcnn/benchmark.sh
+++ b/python/examples/pipeline/PaddleDetection/faster_rcnn/benchmark.sh
@@ -1,36 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="faster_rcnn_r50_fpn_1x_coco"
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="det-FasterRCNN"
+
 # HTTP
-ps -ef | grep web_service | awk '{print $2}' | xargs kill -9 
+#ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
 sleep 3
-python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$modelname
-for thread_num in 1 
+
+echo "Starting HTTP Clients..."
+# Start a client in each thread, tesing the case of multiple threads.
+for thread_num in 1 2 4 8 12 16
 do
-  for batch_size in 1 
+  for batch_size in 1
   do
-    echo "#----FasterRCNN thread num: $thread_num batch size: $batch_size mode:http ----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${gpu_id} --query-compute-apps=used_memory --format=csv -lms 100 > gpu_use.log 2>&1 &
-    nvidia-smi --id=${gpu_id} --query-gpu=utilization.gpu --format=csv -lms 100 > gpu_utilization.log 2>&1 &
+    echo "----${modelname} thread num: ${thread_num} batch size: ${batch_size} mode:http ----" >>profile_log_$modelname
+    # Start one web service, If you start the service yourself, you can ignore it here.
+    #python3 web_service.py >web.log 2>&1 &
+    #sleep 3
+
+    # --id is the serial number of the GPU card, Must be the same as the gpu id used by the server.
+    nvidia-smi --id=3 --query-gpu=memory.used --format=csv -lms 1000 > gpu_use.log 2>&1 &
+    nvidia-smi --id=3 --query-gpu=utilization.gpu --format=csv -lms 1000 > gpu_utilization.log 2>&1 &
     echo "import psutil\ncpu_utilization=psutil.cpu_percent(1,False)\nprint('CPU_UTILIZATION:', cpu_utilization)\n" > cpu_utilization.py
-    python3 benchmark.py run http $thread_num $batch_size
-    python3 cpu_utilization.py >>profile_log_$modelname
-    python3 -m paddle_serving_server_gpu.profiler >>profile_log_$modelname
-    ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
-    ps -ef | grep nvidia-smi | awk '{print $2}' | xargs kill -9
-    python3 benchmark.py dump benchmark.log benchmark.tmp
-    mv benchmark.tmp benchmark.log
-    awk 'BEGIN {max = 0} {if(NR>1){if ($modelname > max) max=$modelname}} END {print "GPU_MEM:", max}' gpu_use.log >> profile_log_$modelname
-    awk 'BEGIN {max = 0} {if(NR>1){if ($modelname > max) max=$modelname}} END {print "GPU_UTIL:", max}' gpu_utilization.log >> profile_log_$modelname
-    cat benchmark.log >> profile_log_$modelname
-    python3 -m paddle_serving_server_gpu.parse_profile --benchmark_cfg $benchmark_config_filename --benchmark_log profile_log_$modelname 
-    #rm -rf gpu_use.log gpu_utilization.log
+    # Start http client
+    python3 benchmark.py run http $thread_num $batch_size > profile 2>&1
+
+    # Collect CPU metrics, Filter data that is zero momentarily, Record the maximum value of GPU memory and the average value of GPU utilization
+    python3 cpu_utilization.py >> profile_log_$modelname
+    grep -av '^0 %' gpu_utilization.log > gpu_utilization.log.tmp
+    awk 'BEGIN {max = 0} {if(NR>1){if ($modelname > max) max=$modelname}} END {print "MAX_GPU_MEMORY:", max}' gpu_use.log >> profile_log_$modelname
+    awk -F' ' '{sum+=$1} END {print "GPU_UTILIZATION:", sum/NR, sum, NR }' gpu_utilization.log.tmp >> profile_log_$modelname
+
+    # Show profiles
+    python3 ../../../util/show_profile.py profile $thread_num >> profile_log_$modelname
+    tail -n 8 profile >> profile_log_$modelname
+    echo '' >> profile_log_$modelname
   done
 done
+
+# Kill all nvidia-smi background task.
+pkill nvidia-smi
diff --git a/python/examples/pipeline/PaddleDetection/ppyolo_mbv3/benchmark.py b/python/examples/pipeline/PaddleDetection/ppyolo_mbv3/benchmark.py
index 714acb12ef8451b967d44b05524791e07ad92056..a23f64314ef448f2617f92ab40f94f75cc6e707f 100644
--- a/python/examples/pipeline/PaddleDetection/ppyolo_mbv3/benchmark.py
+++ b/python/examples/pipeline/PaddleDetection/ppyolo_mbv3/benchmark.py
@@ -36,7 +36,7 @@ def gen_yml(device, gpu_id):
     config["dag"]["tracer"] = {"interval_s": 30}
     if device == "gpu":
         config["op"]["ppyolo_mbv3"]["local_service_conf"]["device_type"] = 1
-        config["op"]["ppyolo_mbv3"]["local_service_conf"]["devices"] = gpu_id        
+        config["op"]["ppyolo_mbv3"]["local_service_conf"]["devices"] = gpu_id
     with open("config2.yml", "w") as fout: 
         yaml.dump(config, fout, default_flow_style=False)
 
@@ -46,23 +46,43 @@ def run_http(idx, batch_size):
     with open(os.path.join(".", "000000570688.jpg"), 'rb') as file:
         image_data1 = file.read()
     image = cv2_to_base64(image_data1)
-
+    latency_list = []
     start = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         data = {"key": [], "value": []}
         for j in range(batch_size):
             data["key"].append("image_" + str(j))
             data["value"].append(image)
         r = requests.post(url=url, data=json.dumps(data))
+        l_end = time.time()
+        total_num += 1
         end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
         if end - start > 70:
-            print("70s end")
+            #print("70s end")
             break
-    return [[end - start]]
+    return [[end - start], latency_list, [total_num]]
 
 def multithread_http(thread, batch_size):
     multi_thread_runner = MultiThreadRunner()
-    result = multi_thread_runner.run(run_http , thread, batch_size)
+    start = time.time()
+    result = multi_thread_runner.run(run_http, thread, batch_size)
+    end = time.time()
+    total_cost = end - start
+    avg_cost = 0
+    total_number = 0
+    for i in range(thread):
+        avg_cost += result[0][i]
+        total_number += result[2][i]
+    avg_cost = avg_cost / thread
+    print("Total cost: {}s".format(total_cost))
+    print("Each thread cost: {}s. ".format(avg_cost))
+    print("Total count: {}. ".format(total_number))
+    print("AVG QPS: {} samples/s".format(batch_size * total_number /
+                                         total_cost))
+    show_latency(result[1])
 
 def run_rpc(thread, batch_size):
     pass
diff --git a/python/examples/pipeline/PaddleDetection/ppyolo_mbv3/benchmark.sh b/python/examples/pipeline/PaddleDetection/ppyolo_mbv3/benchmark.sh
index 1a7cb64429b7f20f1bd01f191518f53b1f82e40c..7237a52ab73630141785f47a143d6931f75d0c17 100644
--- a/python/examples/pipeline/PaddleDetection/ppyolo_mbv3/benchmark.sh
+++ b/python/examples/pipeline/PaddleDetection/ppyolo_mbv3/benchmark.sh
@@ -1,36 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="ppyolo_mbv3_large"
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="det-PPYoloMbv3"
 
 # HTTP
-ps -ef | grep web_service | awk '{print $2}' | xargs kill -9 
+#ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
 sleep 3
-python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$modelname
-for thread_num in 1 
+
+echo "Starting HTTP Clients..."
+# Start a client in each thread, tesing the case of multiple threads.
+for thread_num in 1 2 4 8 12 16
 do
-  for batch_size in 1 
+  for batch_size in 1
   do
-    echo "#----PPyolo thread num: $thread_num batch size: $batch_size mode:http ----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${gpu_id} --query-compute-apps=used_memory --format=csv -lms 100 > gpu_use.log 2>&1 &
-    nvidia-smi --id=${gpu_id} --query-gpu=utilization.gpu --format=csv -lms 100 > gpu_utilization.log 2>&1 &
+    echo "----${modelname} thread num: ${thread_num} batch size: ${batch_size} mode:http ----" >>profile_log_$modelname
+    # Start one web service, If you start the service yourself, you can ignore it here.
+    #python3 web_service.py >web.log 2>&1 &
+    #sleep 3
+
+    # --id is the serial number of the GPU card, Must be the same as the gpu id used by the server.
+    nvidia-smi --id=3 --query-gpu=memory.used --format=csv -lms 1000 > gpu_use.log 2>&1 &
+    nvidia-smi --id=3 --query-gpu=utilization.gpu --format=csv -lms 1000 > gpu_utilization.log 2>&1 &
     echo "import psutil\ncpu_utilization=psutil.cpu_percent(1,False)\nprint('CPU_UTILIZATION:', cpu_utilization)\n" > cpu_utilization.py
-    python3 benchmark.py run http $thread_num $batch_size
-    python3 cpu_utilization.py >>profile_log_$modelname
-    python3 -m paddle_serving_server_gpu.profiler >>profile_log_$modelname
-    ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
-    python3 benchmark.py dump benchmark.log benchmark.tmp
-    mv benchmark.tmp benchmark.log
-    awk 'BEGIN {max = 0} {if(NR>1){if ($modelname > max) max=$modelname}} END {print "GPU_MEM:", max}' gpu_use.log >> profile_log_$modelname
-    awk 'BEGIN {max = 0} {if(NR>1){if ($modelname > max) max=$modelname}} END {print "GPU_UTIL:", max}' gpu_utilization.log >> profile_log_$modelname
-    cat benchmark.log >> profile_log_$modelname
-    python3 -m paddle_serving_server_gpu.parse_profile --benchmark_cfg $benchmark_config_filename --benchmark_log profile_log_$modelname
-    #rm -rf gpu_use.log gpu_utilization.log
+    # Start http client
+    python3 benchmark.py run http $thread_num $batch_size > profile 2>&1
+
+    # Collect CPU metrics, Filter data that is zero momentarily, Record the maximum value of GPU memory and the average value of GPU utilization
+    python3 cpu_utilization.py >> profile_log_$modelname
+    grep -av '^0 %' gpu_utilization.log > gpu_utilization.log.tmp
+    awk 'BEGIN {max = 0} {if(NR>1){if ($modelname > max) max=$modelname}} END {print "MAX_GPU_MEMORY:", max}' gpu_use.log >> profile_log_$modelname
+    awk -F' ' '{sum+=$1} END {print "GPU_UTILIZATION:", sum/NR, sum, NR }' gpu_utilization.log.tmp >> profile_log_$modelname
+
+    # Show profiles
+    python3 ../../../util/show_profile.py profile $thread_num >> profile_log_$modelname
+    tail -n 8 profile >> profile_log_$modelname
+    echo '' >> profile_log_$modelname
   done
 done
+
+# Kill all nvidia-smi background task.
+pkill nvidia-smi
diff --git a/python/examples/pipeline/PaddleDetection/yolov3/benchmark.py b/python/examples/pipeline/PaddleDetection/yolov3/benchmark.py
index 2918c111e4a90b1c460780f6ecb7cecc6df09335..ae9c5a8fb25f56eebe3c3893a4a4d251f21e5b61 100644
--- a/python/examples/pipeline/PaddleDetection/yolov3/benchmark.py
+++ b/python/examples/pipeline/PaddleDetection/yolov3/benchmark.py
@@ -35,8 +35,8 @@ def gen_yml(device, gpu_id):
     fin.close()
     config["dag"]["tracer"] = {"interval_s": 30}
     if device == "gpu":
-        config["op"]["faster_rcnn"]["local_service_conf"]["device_type"] = 1
-        config["op"]["faster_rcnn"]["local_service_conf"]["devices"] = gpu_id        
+        config["op"]["yolov3"]["local_service_conf"]["device_type"] = 1
+        config["op"]["yolov3"]["local_service_conf"]["devices"] = gpu_id        
     with open("config2.yml", "w") as fout: 
         yaml.dump(config, fout, default_flow_style=False)
 
@@ -46,23 +46,43 @@ def run_http(idx, batch_size):
     with open(os.path.join(".", "000000570688.jpg"), 'rb') as file:
         image_data1 = file.read()
     image = cv2_to_base64(image_data1)
-
+    latency_list = []
     start = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         data = {"key": [], "value": []}
         for j in range(batch_size):
             data["key"].append("image_" + str(j))
             data["value"].append(image)
         r = requests.post(url=url, data=json.dumps(data))
+        l_end = time.time()
+        total_num += 1
         end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
         if end - start > 70:
-            print("70s end")
+            #print("70s end")
             break
-    return [[end - start]]
+    return [[end - start], latency_list, [total_num]]
 
 def multithread_http(thread, batch_size):
     multi_thread_runner = MultiThreadRunner()
-    result = multi_thread_runner.run(run_http , thread, batch_size)
+    start = time.time()
+    result = multi_thread_runner.run(run_http, thread, batch_size)
+    end = time.time()
+    total_cost = end - start
+    avg_cost = 0
+    total_number = 0
+    for i in range(thread):
+        avg_cost += result[0][i]
+        total_number += result[2][i]
+    avg_cost = avg_cost / thread
+    print("Total cost: {}s".format(total_cost))
+    print("Each thread cost: {}s. ".format(avg_cost))
+    print("Total count: {}. ".format(total_number))
+    print("AVG QPS: {} samples/s".format(batch_size * total_number /
+                                         total_cost))
+    show_latency(result[1])
 
 def run_rpc(thread, batch_size):
     pass
diff --git a/python/examples/pipeline/PaddleDetection/yolov3/benchmark.sh b/python/examples/pipeline/PaddleDetection/yolov3/benchmark.sh
index 74ab42bd55cd6b9d26c593d6e13d97ff99b3c4e9..e3ac2f79398feea88a1d547047af6db0691c1581 100644
--- a/python/examples/pipeline/PaddleDetection/yolov3/benchmark.sh
+++ b/python/examples/pipeline/PaddleDetection/yolov3/benchmark.sh
@@ -1,36 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="yolov3_darknet53_270e_coco"
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="det-yolov3"
 
 # HTTP
-ps -ef | grep web_service | awk '{print $2}' | xargs kill -9 
+#ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
 sleep 3
-python3 benchmark.py yaml local_predictor 1 cpu 
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$modelname
-for thread_num in 1 8 16
+
+echo "Starting HTTP Clients..."
+# Start a client in each thread, tesing the case of multiple threads.
+for thread_num in 1 2 4 8 12 16
 do
-  for batch_size in 1 
+  for batch_size in 1
   do
-    echo "#----Yolov3 thread num: $thread_num batch size: $batch_size mode:http ----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${gpu_id} --query-compute-apps=used_memory --format=csv -lms 100 > gpu_use.log 2>&1 &
-    nvidia-smi --id=${gpu_id} --query-gpu=utilization.gpu --format=csv -lms 100 > gpu_utilization.log 2>&1 &
+    echo "----${modelname} thread num: ${thread_num} batch size: ${batch_size} mode:http ----" >>profile_log_$modelname
+    # Start one web service, If you start the service yourself, you can ignore it here.
+    #python3 web_service.py >web.log 2>&1 &
+    #sleep 3
+
+    # --id is the serial number of the GPU card, Must be the same as the gpu id used by the server.
+    nvidia-smi --id=3 --query-gpu=memory.used --format=csv -lms 1000 > gpu_use.log 2>&1 &
+    nvidia-smi --id=3 --query-gpu=utilization.gpu --format=csv -lms 1000 > gpu_utilization.log 2>&1 &
     echo "import psutil\ncpu_utilization=psutil.cpu_percent(1,False)\nprint('CPU_UTILIZATION:', cpu_utilization)\n" > cpu_utilization.py
-    python3 benchmark.py run http $thread_num $batch_size
-    python3 cpu_utilization.py >>profile_log_$modelname
-    python3 -m paddle_serving_server_gpu.profiler >>profile_log_$modelname
-    ps -ef | grep web_service | awk '{print $2}' | xargs kill -9
-    python3 benchmark.py dump benchmark.log benchmark.tmp
-    mv benchmark.tmp benchmark.log
-    awk 'BEGIN {max = 0} {if(NR>1){if ($modelname > max) max=$modelname}} END {print "GPU_MEM:", max}' gpu_use.log >> profile_log_$modelname
-    awk 'BEGIN {max = 0} {if(NR>1){if ($modelname > max) max=$modelname}} END {print "GPU_UTIL:", max}' gpu_utilization.log >> profile_log_$modelname
-    cat benchmark.log >> profile_log_$modelname
-    python3 -m paddle_serving_server_gpu.parse_profile --benchmark_cfg $benchmark_config_filename --benchmark_log profile_log_$modelname
-    #rm -rf gpu_use.log gpu_utilization.log
+    # Start http client
+    python3 benchmark.py run http $thread_num $batch_size > profile 2>&1
+
+    # Collect CPU metrics, Filter data that is zero momentarily, Record the maximum value of GPU memory and the average value of GPU utilization
+    python3 cpu_utilization.py >> profile_log_$modelname
+    grep -av '^0 %' gpu_utilization.log > gpu_utilization.log.tmp
+    awk 'BEGIN {max = 0} {if(NR>1){if ($modelname > max) max=$modelname}} END {print "MAX_GPU_MEMORY:", max}' gpu_use.log >> profile_log_$modelname
+    awk -F' ' '{sum+=$1} END {print "GPU_UTILIZATION:", sum/NR, sum, NR }' gpu_utilization.log.tmp >> profile_log_$modelname
+
+    # Show profiles
+    python3 ../../../util/show_profile.py profile $thread_num >> profile_log_$modelname
+    tail -n 8 profile >> profile_log_$modelname
+    echo '' >> profile_log_$modelname
   done
 done
+
+# Kill all nvidia-smi background task.
+pkill nvidia-smi