From a4d6e1aa363930525a49640846bc5e335b35c3d8 Mon Sep 17 00:00:00 2001
From: bjjwwang <wangjiawei04@baidu.com>
Date: Mon, 17 May 2021 11:22:46 +0000
Subject: [PATCH] fix benchmark sh

---
 .../PaddleClas/DarkNet53/benchmark.py         | 25 +++++++-
 .../PaddleClas/DarkNet53/benchmark.sh         | 64 ++++++++++---------
 .../PaddleClas/HRNet_W18_C/benchmark.py       | 25 +++++++-
 .../PaddleClas/HRNet_W18_C/benchmark.sh       | 64 ++++++++++---------
 .../PaddleClas/MobileNetV1/benchmark.py       | 25 +++++++-
 .../PaddleClas/MobileNetV1/benchmark.sh       | 64 ++++++++++---------
 .../PaddleClas/MobileNetV2/benchmark.py       | 25 +++++++-
 .../PaddleClas/MobileNetV2/benchmark.sh       | 64 ++++++++++---------
 .../MobileNetV3_large_x1_0/benchmark.py       | 25 +++++++-
 .../MobileNetV3_large_x1_0/benchmark.sh       | 63 +++++++++---------
 .../ResNeXt101_vd_64x4d/benchmark.py          | 25 +++++++-
 .../ResNeXt101_vd_64x4d/benchmark.sh          | 64 ++++++++++---------
 .../PaddleClas/ResNet50_vd/benchmark.py       | 25 +++++++-
 .../PaddleClas/ResNet50_vd/benchmark.sh       | 64 ++++++++++---------
 .../PaddleClas/ResNet50_vd_FPGM/benchmark.py  | 25 +++++++-
 .../PaddleClas/ResNet50_vd_FPGM/benchmark.sh  | 64 ++++++++++---------
 .../PaddleClas/ResNet50_vd_KL/benchmark.py    | 25 +++++++-
 .../PaddleClas/ResNet50_vd_KL/benchmark.sh    | 64 ++++++++++---------
 .../PaddleClas/ResNet50_vd_PACT/benchmark.py  | 25 +++++++-
 .../PaddleClas/ResNet50_vd_PACT/benchmark.sh  | 64 ++++++++++---------
 .../PaddleClas/ShuffleNetV2_x1_0/benchmark.py | 25 +++++++-
 .../PaddleClas/ShuffleNetV2_x1_0/benchmark.sh | 64 ++++++++++---------
 22 files changed, 616 insertions(+), 362 deletions(-)

diff --git a/python/examples/pipeline/PaddleClas/DarkNet53/benchmark.py b/python/examples/pipeline/PaddleClas/DarkNet53/benchmark.py
index 10078994..2433b013 100644
--- a/python/examples/pipeline/PaddleClas/DarkNet53/benchmark.py
+++ b/python/examples/pipeline/PaddleClas/DarkNet53/benchmark.py
@@ -53,18 +53,39 @@ def run_http(idx, batch_size):
         keys.append("image_{}".format(i))
         values.append(image)
     data = {"key": keys, "value": values}
+    latency_list = []
     start_time = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         r = requests.post(url=url, data=json.dumps(data))
         print(r.json())
+        l_end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
+        total_num += 1
         if time.time() - start_time > 20:
             break
     end = time.time()
-    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):
     client = PipelineClient()
diff --git a/python/examples/pipeline/PaddleClas/DarkNet53/benchmark.sh b/python/examples/pipeline/PaddleClas/DarkNet53/benchmark.sh
index 7c5a57f5..ee20eccf 100644
--- a/python/examples/pipeline/PaddleClas/DarkNet53/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/DarkNet53/benchmark.sh
@@ -1,42 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="imagenet"
-use_gpu=0
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="clas-DarkNet53"
 
 # 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
-if [ $use_gpu -eq 1 ]; then
-  python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
-else
-  python3 benchmark.py yaml local_predictor 1 cpu
-fi
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$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
   do
-    echo "#----imagenet thread num: $thread_num batch size: $batch_size mode:http use_gpu:$use_gpu----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 resnet50_web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${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/PaddleClas/HRNet_W18_C/benchmark.py b/python/examples/pipeline/PaddleClas/HRNet_W18_C/benchmark.py
index 10078994..2433b013 100644
--- a/python/examples/pipeline/PaddleClas/HRNet_W18_C/benchmark.py
+++ b/python/examples/pipeline/PaddleClas/HRNet_W18_C/benchmark.py
@@ -53,18 +53,39 @@ def run_http(idx, batch_size):
         keys.append("image_{}".format(i))
         values.append(image)
     data = {"key": keys, "value": values}
+    latency_list = []
     start_time = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         r = requests.post(url=url, data=json.dumps(data))
         print(r.json())
+        l_end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
+        total_num += 1
         if time.time() - start_time > 20:
             break
     end = time.time()
-    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):
     client = PipelineClient()
diff --git a/python/examples/pipeline/PaddleClas/HRNet_W18_C/benchmark.sh b/python/examples/pipeline/PaddleClas/HRNet_W18_C/benchmark.sh
index 7c5a57f5..ee20eccf 100644
--- a/python/examples/pipeline/PaddleClas/HRNet_W18_C/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/HRNet_W18_C/benchmark.sh
@@ -1,42 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="imagenet"
-use_gpu=0
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="clas-DarkNet53"
 
 # 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
-if [ $use_gpu -eq 1 ]; then
-  python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
-else
-  python3 benchmark.py yaml local_predictor 1 cpu
-fi
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$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
   do
-    echo "#----imagenet thread num: $thread_num batch size: $batch_size mode:http use_gpu:$use_gpu----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 resnet50_web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${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/PaddleClas/MobileNetV1/benchmark.py b/python/examples/pipeline/PaddleClas/MobileNetV1/benchmark.py
index 10078994..2433b013 100644
--- a/python/examples/pipeline/PaddleClas/MobileNetV1/benchmark.py
+++ b/python/examples/pipeline/PaddleClas/MobileNetV1/benchmark.py
@@ -53,18 +53,39 @@ def run_http(idx, batch_size):
         keys.append("image_{}".format(i))
         values.append(image)
     data = {"key": keys, "value": values}
+    latency_list = []
     start_time = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         r = requests.post(url=url, data=json.dumps(data))
         print(r.json())
+        l_end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
+        total_num += 1
         if time.time() - start_time > 20:
             break
     end = time.time()
-    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):
     client = PipelineClient()
diff --git a/python/examples/pipeline/PaddleClas/MobileNetV1/benchmark.sh b/python/examples/pipeline/PaddleClas/MobileNetV1/benchmark.sh
index 7c5a57f5..ee20eccf 100644
--- a/python/examples/pipeline/PaddleClas/MobileNetV1/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/MobileNetV1/benchmark.sh
@@ -1,42 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="imagenet"
-use_gpu=0
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="clas-DarkNet53"
 
 # 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
-if [ $use_gpu -eq 1 ]; then
-  python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
-else
-  python3 benchmark.py yaml local_predictor 1 cpu
-fi
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$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
   do
-    echo "#----imagenet thread num: $thread_num batch size: $batch_size mode:http use_gpu:$use_gpu----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 resnet50_web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${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/PaddleClas/MobileNetV2/benchmark.py b/python/examples/pipeline/PaddleClas/MobileNetV2/benchmark.py
index 10078994..2433b013 100644
--- a/python/examples/pipeline/PaddleClas/MobileNetV2/benchmark.py
+++ b/python/examples/pipeline/PaddleClas/MobileNetV2/benchmark.py
@@ -53,18 +53,39 @@ def run_http(idx, batch_size):
         keys.append("image_{}".format(i))
         values.append(image)
     data = {"key": keys, "value": values}
+    latency_list = []
     start_time = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         r = requests.post(url=url, data=json.dumps(data))
         print(r.json())
+        l_end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
+        total_num += 1
         if time.time() - start_time > 20:
             break
     end = time.time()
-    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):
     client = PipelineClient()
diff --git a/python/examples/pipeline/PaddleClas/MobileNetV2/benchmark.sh b/python/examples/pipeline/PaddleClas/MobileNetV2/benchmark.sh
index 5ae26eb1..53ede791 100644
--- a/python/examples/pipeline/PaddleClas/MobileNetV2/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/MobileNetV2/benchmark.sh
@@ -1,42 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="imagenet"
-use_gpu=0
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="clas-MobileNetV2"
 
 # 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
-if [ $use_gpu -eq 1 ]; then
-  python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
-else
-  python3 benchmark.py yaml local_predictor 1 cpu
-fi
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$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
   do
-    echo "#----imagenet thread num: $thread_num batch size: $batch_size mode:http use_gpu:$use_gpu----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 resnet50_web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${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/PaddleClas/MobileNetV3_large_x1_0/benchmark.py b/python/examples/pipeline/PaddleClas/MobileNetV3_large_x1_0/benchmark.py
index 10078994..2433b013 100644
--- a/python/examples/pipeline/PaddleClas/MobileNetV3_large_x1_0/benchmark.py
+++ b/python/examples/pipeline/PaddleClas/MobileNetV3_large_x1_0/benchmark.py
@@ -53,18 +53,39 @@ def run_http(idx, batch_size):
         keys.append("image_{}".format(i))
         values.append(image)
     data = {"key": keys, "value": values}
+    latency_list = []
     start_time = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         r = requests.post(url=url, data=json.dumps(data))
         print(r.json())
+        l_end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
+        total_num += 1
         if time.time() - start_time > 20:
             break
     end = time.time()
-    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):
     client = PipelineClient()
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 4c008b4f..ee20eccf 100644
--- a/python/examples/pipeline/PaddleClas/MobileNetV3_large_x1_0/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/MobileNetV3_large_x1_0/benchmark.sh
@@ -1,41 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="imagenet"
-use_gpu=0
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="clas-DarkNet53"
 
 # 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
-if [ $use_gpu -eq 1 ]; then
-  python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
-else
-  python3 benchmark.py yaml local_predictor 1 cpu
-fi
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$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
   do
-    echo "#----imagenet thread num: $thread_num batch size: $batch_size mode:http use_gpu:$use_gpu----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 resnet50_web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${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/PaddleClas/ResNeXt101_vd_64x4d/benchmark.py b/python/examples/pipeline/PaddleClas/ResNeXt101_vd_64x4d/benchmark.py
index 10078994..2433b013 100644
--- a/python/examples/pipeline/PaddleClas/ResNeXt101_vd_64x4d/benchmark.py
+++ b/python/examples/pipeline/PaddleClas/ResNeXt101_vd_64x4d/benchmark.py
@@ -53,18 +53,39 @@ def run_http(idx, batch_size):
         keys.append("image_{}".format(i))
         values.append(image)
     data = {"key": keys, "value": values}
+    latency_list = []
     start_time = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         r = requests.post(url=url, data=json.dumps(data))
         print(r.json())
+        l_end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
+        total_num += 1
         if time.time() - start_time > 20:
             break
     end = time.time()
-    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):
     client = PipelineClient()
diff --git a/python/examples/pipeline/PaddleClas/ResNeXt101_vd_64x4d/benchmark.sh b/python/examples/pipeline/PaddleClas/ResNeXt101_vd_64x4d/benchmark.sh
index 7c5a57f5..ee20eccf 100644
--- a/python/examples/pipeline/PaddleClas/ResNeXt101_vd_64x4d/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/ResNeXt101_vd_64x4d/benchmark.sh
@@ -1,42 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="imagenet"
-use_gpu=0
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="clas-DarkNet53"
 
 # 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
-if [ $use_gpu -eq 1 ]; then
-  python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
-else
-  python3 benchmark.py yaml local_predictor 1 cpu
-fi
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$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
   do
-    echo "#----imagenet thread num: $thread_num batch size: $batch_size mode:http use_gpu:$use_gpu----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 resnet50_web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${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/PaddleClas/ResNet50_vd/benchmark.py b/python/examples/pipeline/PaddleClas/ResNet50_vd/benchmark.py
index 10078994..2433b013 100644
--- a/python/examples/pipeline/PaddleClas/ResNet50_vd/benchmark.py
+++ b/python/examples/pipeline/PaddleClas/ResNet50_vd/benchmark.py
@@ -53,18 +53,39 @@ def run_http(idx, batch_size):
         keys.append("image_{}".format(i))
         values.append(image)
     data = {"key": keys, "value": values}
+    latency_list = []
     start_time = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         r = requests.post(url=url, data=json.dumps(data))
         print(r.json())
+        l_end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
+        total_num += 1
         if time.time() - start_time > 20:
             break
     end = time.time()
-    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):
     client = PipelineClient()
diff --git a/python/examples/pipeline/PaddleClas/ResNet50_vd/benchmark.sh b/python/examples/pipeline/PaddleClas/ResNet50_vd/benchmark.sh
index 7c5a57f5..ee20eccf 100644
--- a/python/examples/pipeline/PaddleClas/ResNet50_vd/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/ResNet50_vd/benchmark.sh
@@ -1,42 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="imagenet"
-use_gpu=0
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="clas-DarkNet53"
 
 # 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
-if [ $use_gpu -eq 1 ]; then
-  python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
-else
-  python3 benchmark.py yaml local_predictor 1 cpu
-fi
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$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
   do
-    echo "#----imagenet thread num: $thread_num batch size: $batch_size mode:http use_gpu:$use_gpu----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 resnet50_web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${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/PaddleClas/ResNet50_vd_FPGM/benchmark.py b/python/examples/pipeline/PaddleClas/ResNet50_vd_FPGM/benchmark.py
index 10078994..2433b013 100644
--- a/python/examples/pipeline/PaddleClas/ResNet50_vd_FPGM/benchmark.py
+++ b/python/examples/pipeline/PaddleClas/ResNet50_vd_FPGM/benchmark.py
@@ -53,18 +53,39 @@ def run_http(idx, batch_size):
         keys.append("image_{}".format(i))
         values.append(image)
     data = {"key": keys, "value": values}
+    latency_list = []
     start_time = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         r = requests.post(url=url, data=json.dumps(data))
         print(r.json())
+        l_end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
+        total_num += 1
         if time.time() - start_time > 20:
             break
     end = time.time()
-    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):
     client = PipelineClient()
diff --git a/python/examples/pipeline/PaddleClas/ResNet50_vd_FPGM/benchmark.sh b/python/examples/pipeline/PaddleClas/ResNet50_vd_FPGM/benchmark.sh
index 7c5a57f5..ee20eccf 100644
--- a/python/examples/pipeline/PaddleClas/ResNet50_vd_FPGM/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/ResNet50_vd_FPGM/benchmark.sh
@@ -1,42 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="imagenet"
-use_gpu=0
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="clas-DarkNet53"
 
 # 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
-if [ $use_gpu -eq 1 ]; then
-  python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
-else
-  python3 benchmark.py yaml local_predictor 1 cpu
-fi
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$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
   do
-    echo "#----imagenet thread num: $thread_num batch size: $batch_size mode:http use_gpu:$use_gpu----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 resnet50_web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${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/PaddleClas/ResNet50_vd_KL/benchmark.py b/python/examples/pipeline/PaddleClas/ResNet50_vd_KL/benchmark.py
index 10078994..2433b013 100644
--- a/python/examples/pipeline/PaddleClas/ResNet50_vd_KL/benchmark.py
+++ b/python/examples/pipeline/PaddleClas/ResNet50_vd_KL/benchmark.py
@@ -53,18 +53,39 @@ def run_http(idx, batch_size):
         keys.append("image_{}".format(i))
         values.append(image)
     data = {"key": keys, "value": values}
+    latency_list = []
     start_time = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         r = requests.post(url=url, data=json.dumps(data))
         print(r.json())
+        l_end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
+        total_num += 1
         if time.time() - start_time > 20:
             break
     end = time.time()
-    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):
     client = PipelineClient()
diff --git a/python/examples/pipeline/PaddleClas/ResNet50_vd_KL/benchmark.sh b/python/examples/pipeline/PaddleClas/ResNet50_vd_KL/benchmark.sh
index 7c5a57f5..ee20eccf 100644
--- a/python/examples/pipeline/PaddleClas/ResNet50_vd_KL/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/ResNet50_vd_KL/benchmark.sh
@@ -1,42 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="imagenet"
-use_gpu=0
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="clas-DarkNet53"
 
 # 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
-if [ $use_gpu -eq 1 ]; then
-  python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
-else
-  python3 benchmark.py yaml local_predictor 1 cpu
-fi
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$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
   do
-    echo "#----imagenet thread num: $thread_num batch size: $batch_size mode:http use_gpu:$use_gpu----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 resnet50_web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${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/PaddleClas/ResNet50_vd_PACT/benchmark.py b/python/examples/pipeline/PaddleClas/ResNet50_vd_PACT/benchmark.py
index 10078994..2433b013 100644
--- a/python/examples/pipeline/PaddleClas/ResNet50_vd_PACT/benchmark.py
+++ b/python/examples/pipeline/PaddleClas/ResNet50_vd_PACT/benchmark.py
@@ -53,18 +53,39 @@ def run_http(idx, batch_size):
         keys.append("image_{}".format(i))
         values.append(image)
     data = {"key": keys, "value": values}
+    latency_list = []
     start_time = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         r = requests.post(url=url, data=json.dumps(data))
         print(r.json())
+        l_end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
+        total_num += 1
         if time.time() - start_time > 20:
             break
     end = time.time()
-    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):
     client = PipelineClient()
diff --git a/python/examples/pipeline/PaddleClas/ResNet50_vd_PACT/benchmark.sh b/python/examples/pipeline/PaddleClas/ResNet50_vd_PACT/benchmark.sh
index 7c5a57f5..ee20eccf 100644
--- a/python/examples/pipeline/PaddleClas/ResNet50_vd_PACT/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/ResNet50_vd_PACT/benchmark.sh
@@ -1,42 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="imagenet"
-use_gpu=0
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="clas-DarkNet53"
 
 # 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
-if [ $use_gpu -eq 1 ]; then
-  python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
-else
-  python3 benchmark.py yaml local_predictor 1 cpu
-fi
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$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
   do
-    echo "#----imagenet thread num: $thread_num batch size: $batch_size mode:http use_gpu:$use_gpu----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 resnet50_web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${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/PaddleClas/ShuffleNetV2_x1_0/benchmark.py b/python/examples/pipeline/PaddleClas/ShuffleNetV2_x1_0/benchmark.py
index 10078994..2433b013 100644
--- a/python/examples/pipeline/PaddleClas/ShuffleNetV2_x1_0/benchmark.py
+++ b/python/examples/pipeline/PaddleClas/ShuffleNetV2_x1_0/benchmark.py
@@ -53,18 +53,39 @@ def run_http(idx, batch_size):
         keys.append("image_{}".format(i))
         values.append(image)
     data = {"key": keys, "value": values}
+    latency_list = []
     start_time = time.time()
+    total_num = 0
     while True:
+        l_start = time.time()
         r = requests.post(url=url, data=json.dumps(data))
         print(r.json())
+        l_end = time.time()
+        latency_list.append(l_end * 1000 - l_start * 1000)
+        total_num += 1
         if time.time() - start_time > 20:
             break
     end = time.time()
-    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):
     client = PipelineClient()
diff --git a/python/examples/pipeline/PaddleClas/ShuffleNetV2_x1_0/benchmark.sh b/python/examples/pipeline/PaddleClas/ShuffleNetV2_x1_0/benchmark.sh
index 7c5a57f5..ee20eccf 100644
--- a/python/examples/pipeline/PaddleClas/ShuffleNetV2_x1_0/benchmark.sh
+++ b/python/examples/pipeline/PaddleClas/ShuffleNetV2_x1_0/benchmark.sh
@@ -1,42 +1,44 @@
 export FLAGS_profile_pipeline=1
-alias python3="python3.7"
-modelname="imagenet"
-use_gpu=0
-gpu_id="0"
-benchmark_config_filename="benchmark_config.yaml"
+alias python3="python3.6"
+modelname="clas-DarkNet53"
 
 # 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
-if [ $use_gpu -eq 1 ]; then
-  python3 benchmark.py yaml local_predictor 1 gpu $gpu_id
-else
-  python3 benchmark.py yaml local_predictor 1 cpu
-fi
+# Create yaml，If you already have the config.yaml, ignore it.
+#python3 benchmark.py yaml local_predictor 1 gpu
 rm -rf profile_log_$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
   do
-    echo "#----imagenet thread num: $thread_num batch size: $batch_size mode:http use_gpu:$use_gpu----" >>profile_log_$modelname
-    rm -rf PipelineServingLogs
-    rm -rf cpu_utilization.py
-    python3 resnet50_web_service.py >web.log 2>&1 &
-    sleep 3
-    nvidia-smi --id=${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
-- 
GitLab