diff --git a/core/cube/cube-api/src/cube_cli.cpp b/core/cube/cube-api/src/cube_cli.cpp
index eee4b0c31ad83ca69d242e81bae3ce4ecfb5bf1a..f43f3deacf21db2f777012539fcd83a8fb70002b 100644
--- a/core/cube/cube-api/src/cube_cli.cpp
+++ b/core/cube/cube-api/src/cube_cli.cpp
@@ -31,8 +31,9 @@ DEFINE_bool(print_output, false, "print output flag");
 DEFINE_int32(thread_num, 1, "thread num");
 std::atomic<int> g_concurrency(0);
 
-std::vector<uint64_t> time_list;
+std::vector<std::vector<uint64_t>> time_list;
 std::vector<uint64_t> request_list;
+int turns = 1000000 / FLAGS_batch;
 
 namespace {
 inline uint64_t time_diff(const struct timeval& start_time,
@@ -97,7 +98,7 @@ int run(int argc, char** argv, int thread_id) {
   while (g_concurrency.load() >= FLAGS_thread_num) {
   }
   g_concurrency++;
-
+  time_list[thread_id].resize(turns);
   while (index < file_size) {
     // uint64_t key = strtoul(buffer, NULL, 10);
 
@@ -121,47 +122,12 @@ int run(int argc, char** argv, int thread_id) {
       }
       ++seek_counter;
       uint64_t seek_cost = time_diff(seek_start, seek_end);
-      seek_cost_total += seek_cost;
-      if (seek_cost > seek_cost_max) {
-        seek_cost_max = seek_cost;
-      }
-      if (seek_cost < seek_cost_min) {
-        seek_cost_min = seek_cost;
-      }
+      time_list[thread_id][request - 1] = seek_cost;
 
       keys.clear();
       values.clear();
     }
   }
-  /*
-    if (keys.size() > 0) {
-      int ret = 0;
-      values.resize(keys.size());
-      TIME_FLAG(seek_start);
-      ret = cube->seek(FLAGS_dict, keys, &values);
-      TIME_FLAG(seek_end);
-      if (ret != 0) {
-        LOG(WARNING) << "cube seek failed";
-      } else if (FLAGS_print_output) {
-        for (size_t i = 0; i < keys.size(); ++i) {
-          fprintf(stdout,
-                  "key:%lu value:%s\n",
-                  keys[i],
-                  string_to_hex(values[i].buff).c_str());
-        }
-      }
-
-      ++seek_counter;
-      uint64_t seek_cost = time_diff(seek_start, seek_end);
-      seek_cost_total += seek_cost;
-      if (seek_cost > seek_cost_max) {
-        seek_cost_max = seek_cost;
-      }
-      if (seek_cost < seek_cost_min) {
-        seek_cost_min = seek_cost;
-      }
-    }
-  */
   g_concurrency--;
 
   // fclose(key_file);
@@ -171,12 +137,6 @@ int run(int argc, char** argv, int thread_id) {
     LOG(WARNING) << "destroy cube api failed err=" << ret;
   }
 
-  uint64_t seek_cost_avg = seek_cost_total / seek_counter;
-  LOG(INFO) << "seek cost avg = " << seek_cost_avg;
-  LOG(INFO) << "seek cost max = " << seek_cost_max;
-  LOG(INFO) << "seek cost min = " << seek_cost_min;
-
-  time_list[thread_id] = seek_cost_avg;
   request_list[thread_id] = request;
 
   return 0;
@@ -188,6 +148,7 @@ int run_m(int argc, char** argv) {
   request_list.resize(thread_num);
   time_list.resize(thread_num);
   std::vector<std::thread*> thread_pool;
+  TIME_FLAG(main_start);
   for (int i = 0; i < thread_num; i++) {
     thread_pool.push_back(new std::thread(run, argc, argv, i));
   }
@@ -195,27 +156,33 @@ int run_m(int argc, char** argv) {
     thread_pool[i]->join();
     delete thread_pool[i];
   }
+  TIME_FLAG(main_end);
   uint64_t sum_time = 0;
   uint64_t max_time = 0;
   uint64_t min_time = 1000000;
   uint64_t request_num = 0;
   for (int i = 0; i < thread_num; i++) {
-    sum_time += time_list[i];
-    if (time_list[i] > max_time) {
-      max_time = time_list[i];
-    }
-    if (time_list[i] < min_time) {
-      min_time = time_list[i];
+    for (int j = 0; j < request_list[i]; j++) {
+      sum_time += time_list[i][j];
+      if (time_list[i][j] > max_time) {
+        max_time = time_list[i][j];
+      }
+      if (time_list[i][j] < min_time) {
+        min_time = time_list[i][j];
+      }
     }
     request_num += request_list[i];
   }
-  uint64_t mean_time = sum_time / thread_num;
-  LOG(INFO) << thread_num << " thread seek cost"
-            << " avg = " << std::to_string(mean_time)
-            << " max = " << std::to_string(max_time)
-            << " min = " << std::to_string(min_time);
-  LOG(INFO) << " total_request = " << std::to_string(request_num) << " speed = "
-            << std::to_string(1000000 * thread_num / mean_time)  // mean_time us
+  uint64_t mean_time = sum_time / (thread_num * turns);
+  uint64_t main_time = time_diff(main_start, main_end);
+  LOG(INFO) << "\n"
+            << thread_num << " thread seek cost"
+            << "\navg = " << std::to_string(mean_time)
+            << "\nmax = " << std::to_string(max_time)
+            << "\nmin = " << std::to_string(min_time);
+  LOG(INFO) << "\ntotal_request = " << std::to_string(request_num)
+            << "\nspeed = " << std::to_string(request_num * 1000000 /
+                                              main_time)  // mean_time us
             << " query per second";
   return 0;
 }
diff --git a/core/general-server/op/general_dist_kv_infer_op.cpp b/core/general-server/op/general_dist_kv_infer_op.cpp
index 9c6c70352b5387fab95acd16cdf79aa2b46f6122..adaa6cbc1818fc5300faf662d98ad47c9af4c468 100644
--- a/core/general-server/op/general_dist_kv_infer_op.cpp
+++ b/core/general-server/op/general_dist_kv_infer_op.cpp
@@ -90,6 +90,9 @@ int GeneralDistKVInferOp::inference() {
               keys.begin() + key_idx);
     key_idx += dataptr_size_pairs[i].second;
   }
+  Timer timeline;
+  int64_t cube_start = timeline.TimeStampUS();
+  timeline.Start();
   rec::mcube::CubeAPI *cube = rec::mcube::CubeAPI::instance();
   std::vector<std::string> table_names = cube->get_table_names();
   if (table_names.size() == 0) {
@@ -97,7 +100,7 @@ int GeneralDistKVInferOp::inference() {
     return -1;
   }
   int ret = cube->seek(table_names[0], keys, &values);
-
+  int64_t cube_end = timeline.TimeStampUS();
   if (values.size() != keys.size() || values[0].buff.size() == 0) {
     LOG(ERROR) << "cube value return null";
   }
@@ -153,9 +156,7 @@ int GeneralDistKVInferOp::inference() {
 
   VLOG(2) << "infer batch size: " << batch_size;
 
-  Timer timeline;
   int64_t start = timeline.TimeStampUS();
-  timeline.Start();
 
   if (InferManager::instance().infer(
           engine_name().c_str(), &infer_in, out, batch_size)) {
@@ -165,6 +166,8 @@ int GeneralDistKVInferOp::inference() {
 
   int64_t end = timeline.TimeStampUS();
   CopyBlobInfo(input_blob, output_blob);
+  AddBlobInfo(output_blob, cube_start);
+  AddBlobInfo(output_blob, cube_end);
   AddBlobInfo(output_blob, start);
   AddBlobInfo(output_blob, end);
   return 0;
diff --git a/python/examples/bert/benchmark.sh b/python/examples/bert/benchmark.sh
index 96c9cd97d15bf1133210feb0bcabf95fdcde9d37..09e9e1bc23b81f118a22a14ffc51fa2fd5a951d4 100644
--- a/python/examples/bert/benchmark.sh
+++ b/python/examples/bert/benchmark.sh
@@ -1,41 +1,52 @@
-rm profile_log
+rm profile_log*
 export CUDA_VISIBLE_DEVICES=0,1,2,3
 export FLAGS_profile_server=1
 export FLAGS_profile_client=1
 export FLAGS_serving_latency=1
-python3 -m paddle_serving_server_gpu.serve --model $1 --port 9292 --thread 4 --gpu_ids 0,1,2,3 --mem_optim False --ir_optim True 2> elog > stdlog &
-hostname=`echo $(hostname)|awk -F '.baidu.com' '{print $1}'`
-sleep 5
+
 gpu_id=0
+#save cpu and gpu utilization log
+if [ -d utilization ];then
+    rm -rf utilization
+else
+    mkdir utilization
+fi
+#start server
+$PYTHONROOT/bin/python3 -m paddle_serving_server_gpu.serve --model $1 --port 9292 --thread 4 --gpu_ids 0,1,2,3 --mem_optim  --ir_optim >  elog  2>&1 &
+sleep 5
 
 #warm up
-python3 benchmark.py --thread 8 --batch_size 1 --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
-
-for thread_num in 4 8 16
+$PYTHONROOT/bin/python3 benchmark.py --thread 4 --batch_size 1 --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+echo -e "import psutil\ncpu_utilization=psutil.cpu_percent(1,False)\nprint('CPU_UTILIZATION:', cpu_utilization)\n" > cpu_utilization.py
+for thread_num in 1 4 8 16
 do
-for batch_size in 1 4 16 64 256
+for batch_size in 1 4 16 64
 do
     job_bt=`date '+%Y%m%d%H%M%S'`
-    nvidia-smi --id=$gpu_id --query-compute-apps=used_memory --format=csv -lms 100 > gpu_use.log 2>&1 &
+    nvidia-smi --id=0 --query-compute-apps=used_memory --format=csv -lms 100 > gpu_use.log 2>&1 &
+    nvidia-smi --id=0 --query-gpu=utilization.gpu --format=csv -lms 100 > gpu_utilization.log 2>&1 &
     gpu_memory_pid=$!
-    python3 benchmark.py --thread $thread_num --batch_size $batch_size --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+    $PYTHONROOT/bin/python3 benchmark.py --thread $thread_num --batch_size $batch_size --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
     kill ${gpu_memory_pid}
+    kill `ps -ef|grep used_memory|awk '{print $2}'`
     echo "model_name:" $1
     echo "thread_num:" $thread_num
     echo "batch_size:" $batch_size
     echo "=================Done===================="
     echo "model_name:$1" >> profile_log_$1
     echo "batch_size:$batch_size" >> profile_log_$1
+    $PYTHONROOT/bin/python3 cpu_utilization.py >> profile_log_$1
     job_et=`date '+%Y%m%d%H%M%S'`
-    awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "MAX_GPU_MEMORY_USE:", max}' gpu_use.log >> profile_log_$1
-    monquery -n ${hostname} -i GPU_AVERAGE_UTILIZATION -s $job_bt -e $job_et -d 10 > gpu_log_file_${job_bt}
-    monquery -n ${hostname} -i CPU_USER -s $job_bt -e $job_et  -d 10 > cpu_log_file_${job_bt}
-    cpu_num=$(cat /proc/cpuinfo | grep processor | wc -l)
-    gpu_num=$(nvidia-smi -L|wc -l)
-    python ../util/show_profile.py profile $thread_num >> profile_log_$1
+    awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "MAX_GPU_MEMORY:", max}' gpu_use.log >> profile_log_$1
+    awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "GPU_UTILIZATION:", max}' gpu_utilization.log >> profile_log_$1
+    rm -rf gpu_use.log gpu_utilization.log
+    $PYTHONROOT/bin/python3 ../util/show_profile.py profile $thread_num >> profile_log_$1
     tail -n 8 profile >> profile_log_$1
     echo "" >> profile_log_$1
 done
 done
 
+#Divided log
+awk 'BEGIN{RS="\n\n"}{i++}{print > "bert_log_"i}' profile_log_$1
+mkdir bert_log && mv bert_log_* bert_log
 ps -ef|grep 'serving'|grep -v grep|cut -c 9-15 | xargs kill -9
diff --git a/python/examples/criteo_ctr_with_cube/benchmark.py b/python/examples/criteo_ctr_with_cube/benchmark.py
index a850d244b0a5a1a01e98a6207fa9674b6ea0af1a..324eb18e214237cdb0d228fc6b57c8efd3665cc9 100755
--- a/python/examples/criteo_ctr_with_cube/benchmark.py
+++ b/python/examples/criteo_ctr_with_cube/benchmark.py
@@ -24,6 +24,7 @@ from paddle_serving_client.utils import MultiThreadRunner
 from paddle_serving_client.utils import benchmark_args
 from paddle_serving_client.metric import auc
 
+py_version = sys.version_info[0]
 args = benchmark_args()
 
 
@@ -49,7 +50,10 @@ def single_func(idx, resource):
             if args.batch_size > 0:
                 feed_batch = []
                 for bi in range(args.batch_size):
-                    data = reader().next()
+                    if py_version == 2:
+                        data = reader().next()
+                    else:
+                        data = reader().__next__()
                     feed_dict = {}
                     feed_dict['dense_input'] = data[0][0]
                     for i in range(1, 27):
@@ -71,14 +75,17 @@ if __name__ == '__main__':
     multi_thread_runner = MultiThreadRunner()
     endpoint_list = ["127.0.0.1:9292"]
     #result = single_func(0, {"endpoint": endpoint_list})
+    start = time.time()
     result = multi_thread_runner.run(single_func, args.thread,
                                      {"endpoint": endpoint_list})
-    print(result)
+    end = time.time()
+    total_cost = end - start
     avg_cost = 0
     qps = 0
     for i in range(args.thread):
         avg_cost += result[0][i * 2 + 0]
         qps += result[0][i * 2 + 1]
     avg_cost = avg_cost / args.thread
+    print("total cost: {}".format(total_cost))
     print("average total cost {} s.".format(avg_cost))
     print("qps {} ins/s".format(qps))
diff --git a/python/examples/criteo_ctr_with_cube/benchmark.sh b/python/examples/criteo_ctr_with_cube/benchmark.sh
index 35b19b637d9e8dec10fd3b59224c5c17e3ba5f53..21daf9331ec4a7ba98ac73fc4570b024681aa06a 100755
--- a/python/examples/criteo_ctr_with_cube/benchmark.sh
+++ b/python/examples/criteo_ctr_with_cube/benchmark.sh
@@ -1,9 +1,23 @@
 rm profile_log
 export FLAGS_profile_client=1
 export FLAGS_profile_server=1
-for thread_num in 1 2 4 8 16
+
+wget https://paddle-serving.bj.bcebos.com/unittest/ctr_cube_unittest.tar.gz --no-check-certificate
+tar xf ctr_cube_unittest.tar.gz
+mv models/ctr_client_conf ./
+mv models/ctr_serving_model_kv ./
+mv models/data ./cube/
+
+wget https://paddle-serving.bj.bcebos.com/others/cube_app.tar.gz --no-check-certificate
+tar xf cube_app.tar.gz
+mv cube_app/cube* ./cube/
+sh cube_prepare.sh &
+
+python test_server.py ctr_serving_model_kv > serving_log 2>&1 &
+
+for thread_num in 1 4 16
 do
-for batch_size in 1 4 16 64 256
+for batch_size in 1 4 16 64
 do
     $PYTHONROOT/bin/python benchmark.py --thread $thread_num --batch_size $batch_size --model serving_client_conf/serving_client_conf.prototxt --request rpc > profile 2>&1
     echo "batch size : $batch_size"
@@ -11,6 +25,8 @@ do
     echo "========================================"
     echo "batch size : $batch_size" >> profile_log
     $PYTHONROOT/bin/python ../util/show_profile.py profile $thread_num >> profile_log
-    tail -n 2 profile >> profile_log
+    tail -n 3 profile >> profile_log
 done
 done
+
+ps -ef|grep 'serving'|grep -v grep|cut -c 9-15 | xargs kill -9
diff --git a/python/examples/criteo_ctr_with_cube/benchmark_cube.sh b/python/examples/criteo_ctr_with_cube/benchmark_cube.sh
new file mode 100755
index 0000000000000000000000000000000000000000..e4a2cb3681642312f9c7fe8199cc4d7e68f5d9bf
--- /dev/null
+++ b/python/examples/criteo_ctr_with_cube/benchmark_cube.sh
@@ -0,0 +1,33 @@
+rm profile_log
+
+wget https://paddle-serving.bj.bcebos.com/unittest/ctr_cube_unittest.tar.gz --no-check-certificate
+tar xf ctr_cube_unittest.tar.gz
+mv models/ctr_client_conf ./
+mv models/ctr_serving_model_kv ./
+mv models/data ./cube/
+
+wget https://paddle-serving.bj.bcebos.com/others/cube_app.tar.gz --no-check-certificate
+tar xf cube_app.tar.gz
+mv cube_app/cube* ./cube/
+sh cube_prepare.sh &
+
+cp ../../../build_server/core/cube/cube-api/cube-cli .
+python gen_key.py
+
+for thread_num in 1 4 16 32
+do
+for batch_size in 1000
+do
+    ./cube-cli -config_file ./cube/conf/cube.conf -keys key -dict test_dict -thread_num $thread_num --batch $batch_size > profile 2>&1
+    echo "batch size : $batch_size"
+    echo "thread num : $thread_num"
+    echo "========================================"
+    echo "batch size : $batch_size" >> profile_log
+    echo "thread num : $thread_num" >> profile_log
+    tail -n 7 profile | head -n 4 >> profile_log
+    tail -n 2 profile >> profile_log
+
+done
+done
+
+ps -ef|grep 'cube'|grep -v grep|cut -c 9-15 | xargs kill -9
diff --git a/python/examples/criteo_ctr_with_cube/gen_key.py b/python/examples/criteo_ctr_with_cube/gen_key.py
new file mode 100644
index 0000000000000000000000000000000000000000..115d81701fb2c8b78085c4c88a685dda992f2c27
--- /dev/null
+++ b/python/examples/criteo_ctr_with_cube/gen_key.py
@@ -0,0 +1,20 @@
+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+import random
+
+with open("key", "w") as f:
+    for i in range(1000000):
+        f.write("{}\n".format(random.randint(0, 999999)))
diff --git a/python/examples/criteo_ctr_with_cube/test_server.py b/python/examples/criteo_ctr_with_cube/test_server.py
index 5399ace839a00071c0ed9ce384e5523b68db27fc..479c602910b5afa52b35a66d00316f54905c0741 100755
--- a/python/examples/criteo_ctr_with_cube/test_server.py
+++ b/python/examples/criteo_ctr_with_cube/test_server.py
@@ -33,5 +33,9 @@ server = Server()
 server.set_op_sequence(op_seq_maker.get_op_sequence())
 server.set_num_threads(4)
 server.load_model_config(sys.argv[1])
-server.prepare_server(workdir="work_dir1", port=9292, device="cpu")
+server.prepare_server(
+    workdir="work_dir1",
+    port=9292,
+    device="cpu",
+    cube_conf="./cube/conf/cube.conf")
 server.run_server()
diff --git a/python/examples/criteo_ctr_with_cube/test_server_gpu.py b/python/examples/criteo_ctr_with_cube/test_server_gpu.py
index 382be99bd37a52630d78bb84ef7e53047b018c95..33f74f91c13fca489db70a4d4171ae756355c787 100755
--- a/python/examples/criteo_ctr_with_cube/test_server_gpu.py
+++ b/python/examples/criteo_ctr_with_cube/test_server_gpu.py
@@ -33,5 +33,9 @@ server = Server()
 server.set_op_sequence(op_seq_maker.get_op_sequence())
 server.set_num_threads(4)
 server.load_model_config(sys.argv[1])
-server.prepare_server(workdir="work_dir1", port=9292, device="cpu")
+server.prepare_server(
+    workdir="work_dir1",
+    port=9292,
+    device="cpu",
+    cube_conf="./cube/conf/cube.conf")
 server.run_server()
diff --git a/python/examples/grpc_impl_example/criteo_ctr_with_cube/test_server.py b/python/examples/grpc_impl_example/criteo_ctr_with_cube/test_server.py
index 361d5a59becb7c110907f66d8b651e05e7eb418e..8a3bee4e628ddd0896c1d2facbccbf2ef493df2b 100755
--- a/python/examples/grpc_impl_example/criteo_ctr_with_cube/test_server.py
+++ b/python/examples/grpc_impl_example/criteo_ctr_with_cube/test_server.py
@@ -33,5 +33,9 @@ server = Server()
 server.set_op_sequence(op_seq_maker.get_op_sequence())
 server.set_num_threads(4)
 server.load_model_config(sys.argv[1], sys.argv[2])
-server.prepare_server(workdir="work_dir1", port=9292, device="cpu")
+server.prepare_server(
+    workdir="work_dir1",
+    port=9292,
+    device="cpu",
+    cube_conf="./cube/conf/cube.conf")
 server.run_server()
diff --git a/python/examples/grpc_impl_example/criteo_ctr_with_cube/test_server_gpu.py b/python/examples/grpc_impl_example/criteo_ctr_with_cube/test_server_gpu.py
index 38e1bf82118f6af7cfe7b467003332a5328b2979..343ded248e2ead554cd0235f890ebefc0b09c071 100755
--- a/python/examples/grpc_impl_example/criteo_ctr_with_cube/test_server_gpu.py
+++ b/python/examples/grpc_impl_example/criteo_ctr_with_cube/test_server_gpu.py
@@ -33,5 +33,9 @@ server = Server()
 server.set_op_sequence(op_seq_maker.get_op_sequence())
 server.set_num_threads(4)
 server.load_model_config(sys.argv[1], sys.argv[2])
-server.prepare_server(workdir="work_dir1", port=9292, device="cpu")
+server.prepare_server(
+    workdir="work_dir1",
+    port=9292,
+    device="cpu",
+    cube_conf="./cube/conf/cube.conf")
 server.run_server()
diff --git a/python/examples/imagenet/benchmark.py b/python/examples/imagenet/benchmark.py
index f4a7b083300be727ba81e880c41791bf36bfd6f7..0181b873a36c0e65beff1d03f750b5d78c89aa06 100644
--- a/python/examples/imagenet/benchmark.py
+++ b/python/examples/imagenet/benchmark.py
@@ -24,7 +24,7 @@ import json
 import base64
 from paddle_serving_client import Client
 from paddle_serving_client.utils import MultiThreadRunner
-from paddle_serving_client.utils import benchmark_args
+from paddle_serving_client.utils import benchmark_args, show_latency
 from paddle_serving_app.reader import Sequential, File2Image, Resize
 from paddle_serving_app.reader import CenterCrop, RGB2BGR, Transpose, Div, Normalize
 
@@ -38,7 +38,11 @@ seq_preprocess = Sequential([
 
 def single_func(idx, resource):
     file_list = []
-    turns = 10
+    turns = resource["turns"]
+    latency_flags = False
+    if os.getenv("FLAGS_serving_latency"):
+        latency_flags = True
+        latency_list = []
     for file_name in os.listdir("./image_data/n01440764"):
         file_list.append(file_name)
     img_list = []
@@ -56,6 +60,7 @@ def single_func(idx, resource):
         start = time.time()
         for i in range(turns):
             if args.batch_size >= 1:
+                l_start = time.time()
                 feed_batch = []
                 i_start = time.time()
                 for bi in range(args.batch_size):
@@ -69,6 +74,9 @@ def single_func(idx, resource):
                                  int(round(i_end * 1000000))))
 
                 result = client.predict(feed=feed_batch, fetch=fetch)
+                l_end = time.time()
+                if latency_flags:
+                    latency_list.append(l_end * 1000 - l_start * 1000)
             else:
                 print("unsupport batch size {}".format(args.batch_size))
 
@@ -88,6 +96,8 @@ def single_func(idx, resource):
             r = requests.post(
                 server, data=req, headers={"Content-Type": "application/json"})
     end = time.time()
+    if latency_flags:
+        return [[end - start], latency_list]
     return [[end - start]]
 
 
@@ -96,11 +106,21 @@ if __name__ == '__main__':
     endpoint_list = [
         "127.0.0.1:9292", "127.0.0.1:9293", "127.0.0.1:9294", "127.0.0.1:9295"
     ]
-    result = multi_thread_runner.run(single_func, args.thread,
-                                     {"endpoint": endpoint_list})
+    turns = 100
+    start = time.time()
+    result = multi_thread_runner.run(
+        single_func, args.thread, {"endpoint": endpoint_list,
+                                   "turns": turns})
     #result = single_func(0, {"endpoint": endpoint_list})
+    end = time.time()
+    total_cost = end - start
     avg_cost = 0
     for i in range(args.thread):
         avg_cost += result[0][i]
     avg_cost = avg_cost / args.thread
-    print("average total cost {} s.".format(avg_cost))
+    print("total cost: {}s".format(end - start))
+    print("each thread cost: {}s.".format(avg_cost))
+    print("qps: {}samples/s".format(args.batch_size * args.thread * turns /
+                                    total_cost))
+    if os.getenv("FLAGS_serving_latency"):
+        show_latency(result[1])
diff --git a/python/examples/imagenet/benchmark.sh b/python/examples/imagenet/benchmark.sh
index d7eb89fa9b0b68e5e442d15bdf16f431c91ba94d..620cf2a3d9fe6c292cedecd84dfda0bce42c15d4 100644
--- a/python/examples/imagenet/benchmark.sh
+++ b/python/examples/imagenet/benchmark.sh
@@ -1,28 +1,50 @@
-rm profile_log
+rm profile_log*
 export CUDA_VISIBLE_DEVICES=0,1,2,3
 export FLAGS_profile_server=1
 export FLAGS_profile_client=1
-python -m paddle_serving_server_gpu.serve --model $1 --port 9292 --thread 4 --gpu_ids 0,1,2,3 2> elog > stdlog &
+python -m paddle_serving_server_gpu.serve --model $1 --port 9292 --thread 4 --gpu_ids 0,1,2,3 --mem_optim --ir_optim  2> elog > stdlog &
 
 sleep 5
+gpu_id=0
+#save cpu and gpu utilization log
+if [ -d utilization ];then
+    rm -rf utilization
+else
+    mkdir utilization
+fi
 
 #warm up
-$PYTHONROOT/bin/python benchmark.py --thread 8 --batch_size 1 --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+$PYTHONROOT/bin/python3 benchmark.py --thread 4 --batch_size 1 --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+echo -e "import psutil\ncpu_utilization=psutil.cpu_percent(1,False)\nprint('CPU_UTILIZATION:', cpu_utilization)\n" > cpu_utilization.py
 
-for thread_num in 4 8 16
+for thread_num in 1 4 8 16
 do
 for batch_size in 1 4 16 64
 do
+    job_bt=`date '+%Y%m%d%H%M%S'`
+    nvidia-smi --id=0 --query-compute-apps=used_memory --format=csv -lms 100 > gpu_use.log 2>&1 &
+    nvidia-smi --id=0 --query-gpu=utilization.gpu --format=csv -lms 100 > gpu_utilization.log 2>&1 &
+    gpu_memory_pid=$!
     $PYTHONROOT/bin/python benchmark.py --thread $thread_num --batch_size $batch_size --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+    kill ${gpu_memory_pid}
+    kill `ps -ef|grep used_memory|awk '{print $2}'`
     echo "model name :" $1
     echo "thread num :" $thread_num
     echo "batch size :" $batch_size
     echo "=================Done===================="
     echo "model name :$1" >> profile_log
     echo "batch size :$batch_size" >> profile_log
+    job_et=`date '+%Y%m%d%H%M%S'`
+    awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "MAX_GPU_MEMORY:", max}' gpu_use.log >> profile_log_$1
+    awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "GPU_UTILIZATION:", max}' gpu_utilization.log >> profile_log_$1
+    rm -rf gpu_use.log gpu_utilization.log
     $PYTHONROOT/bin/python ../util/show_profile.py profile $thread_num >> profile_log
     tail -n 8 profile >> profile_log
+    echo "" >> profile_log_$1
 done
 done
 
+#Divided log
+awk 'BEGIN{RS="\n\n"}{i++}{print > "ResNet_log_"i}' profile_log_$1
+mkdir $1_log && mv ResNet_log_* $1_log
 ps -ef|grep 'serving'|grep -v grep|cut -c 9-15 | xargs kill -9
diff --git a/python/examples/imdb/benchmark.sh b/python/examples/imdb/benchmark.sh
index 2b2d91c8192e66f6b3eee19b59fc7f5dc9339aa6..7db9a1086314047930bee32fe8c695c2b71753bf 100644
--- a/python/examples/imdb/benchmark.sh
+++ b/python/examples/imdb/benchmark.sh
@@ -1,19 +1,28 @@
-rm profile_log
-export CUDA_VISIBLE_DEVICES=0,1,2,3
+rm profile_log*
 export FLAGS_profile_server=1
 export FLAGS_profile_client=1
 export FLAGS_serving_latency=1
-python -m paddle_serving_server_gpu.serve --model $1 --port 9292 --thread 4 --gpu_ids 0,1,2,3 --mem_optim --ir_optim  2> elog > stdlog &
+$PYTHONROOT/bin/python3 -m paddle_serving_server.serve --model $1 --port 9292 --thread 4 --mem_optim --ir_optim  2> elog > stdlog &
 hostname=`echo $(hostname)|awk -F '.baidu.com' '{print $1}'`
-
+#save cpu and gpu utilization log
+if [ -d utilization ];then
+    rm -rf utilization
+else
+    mkdir utilization
+fi
 sleep 5
 
-for thread_num in 4 8 16
+
+#warm up
+$PYTHONROOT/bin/python3 benchmark.py --thread 4 --batch_size 1 --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+echo -e "import psutil\ncpu_utilization=psutil.cpu_percent(1,False)\nprint('CPU_UTILIZATION:', cpu_utilization)\n" > cpu_utilization.py
+
+for thread_num in 1 4 8 16
 do
-for batch_size in 1 4 16 64 256
+for batch_size in 1 4 16 64
 do
     job_bt=`date '+%Y%m%d%H%M%S'`
-    python benchmark.py --thread $thread_num --batch_size $batch_size --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+    $PYTHONROOT/bin/python3 benchmark.py --thread $thread_num --batch_size $batch_size --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
     echo "model_name:" $1
     echo "thread_num:" $thread_num
     echo "batch_size:" $batch_size
@@ -21,15 +30,14 @@ do
     echo "model_name:$1" >> profile_log_$1
     echo "batch_size:$batch_size" >> profile_log_$1
     job_et=`date '+%Y%m%d%H%M%S'`
-    awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "MAX_GPU_MEMORY_USE:", max}' gpu_use.log >> profile_log_$1
-    monquery -n ${hostname} -i GPU_AVERAGE_UTILIZATION -s $job_bt -e $job_et -d 10 > gpu_log_file_${job_bt}
-    monquery -n ${hostname} -i CPU_USER -s $job_bt -e $job_et  -d 10 > cpu_log_file_${job_bt}
-    cpu_num=$(cat /proc/cpuinfo | grep processor | wc -l)
-    gpu_num=$(nvidia-smi -L|wc -l)
-    python ../util/show_profile.py profile $thread_num >> profile_log_$1
+    $PYTHONROOT/bin/python3 ../util/show_profile.py profile $thread_num >> profile_log_$1
+    $PYTHONROOT/bin/python3 cpu_utilization.py >> profile_log_$1
     tail -n 8 profile >> profile_log_$1
     echo "" >> profile_log_$1
 done
 done
 
+#Divided log
+awk 'BEGIN{RS="\n\n"}{i++}{print > "imdb_log_"i}' profile_log_$1
+mkdir $1_log && mv imdb_log_* $1_log
 ps -ef|grep 'serving'|grep -v grep|cut -c 9-15 | xargs kill -9
diff --git a/python/paddle_serving_app/reader/image_reader.py b/python/paddle_serving_app/reader/image_reader.py
index 096f46549af137cb04a87e26a3b28c8d42e33daa..cfd50e2514ae796af06d3301fc728bf0f5a88ad0 100644
--- a/python/paddle_serving_app/reader/image_reader.py
+++ b/python/paddle_serving_app/reader/image_reader.py
@@ -677,7 +677,7 @@ class Resize(object):
 
     Args:
         size (sequence or int): Desired output size. If size is a sequence like
-            (h, w), output size will be matched to this. If size is an int,
+            (w, h), output size will be matched to this. If size is an int,
             smaller edge of the image will be matched to this number.
             i.e, if height > width, then image will be rescaled to
             (size * height / width, size)
diff --git a/python/paddle_serving_server/__init__.py b/python/paddle_serving_server/__init__.py
index 875e275c759d9fb1a9ccb6632816418a75a93aec..1e5fd16ed6c153a28cd72422ca3ef7b9177cb079 100644
--- a/python/paddle_serving_server/__init__.py
+++ b/python/paddle_serving_server/__init__.py
@@ -25,6 +25,7 @@ from contextlib import closing
 import collections
 import fcntl
 
+import shutil
 import numpy as np
 import grpc
 from .proto import multi_lang_general_model_service_pb2
@@ -230,7 +231,7 @@ class Server(object):
             infer_service.workflows.extend(["workflow1"])
             self.infer_service_conf.services.extend([infer_service])
 
-    def _prepare_resource(self, workdir):
+    def _prepare_resource(self, workdir, cube_conf):
         self.workdir = workdir
         if self.resource_conf == None:
             with open("{}/{}".format(workdir, self.general_model_config_fn),
@@ -242,6 +243,11 @@ class Server(object):
                     if "dist_kv" in node.name:
                         self.resource_conf.cube_config_path = workdir
                         self.resource_conf.cube_config_file = self.cube_config_fn
+                        if cube_conf == None:
+                            raise ValueError(
+                                "Please set the path of cube.conf while use dist_kv op."
+                            )
+                        shutil.copy(cube_conf, workdir)
                         if "quant" in node.name:
                             self.resource_conf.cube_quant_bits = 8
             self.resource_conf.model_toolkit_path = workdir
@@ -366,7 +372,11 @@ class Server(object):
         os.chdir(self.cur_path)
         self.bin_path = self.server_path + "/serving"
 
-    def prepare_server(self, workdir=None, port=9292, device="cpu"):
+    def prepare_server(self,
+                       workdir=None,
+                       port=9292,
+                       device="cpu",
+                       cube_conf=None):
         if workdir == None:
             workdir = "./tmp"
             os.system("mkdir {}".format(workdir))
@@ -377,7 +387,7 @@ class Server(object):
         if not self.port_is_available(port):
             raise SystemExit("Port {} is already used".format(port))
         self.set_port(port)
-        self._prepare_resource(workdir)
+        self._prepare_resource(workdir, cube_conf)
         self._prepare_engine(self.model_config_paths, device)
         self._prepare_infer_service(port)
         self.workdir = workdir
@@ -645,7 +655,11 @@ class MultiLangServer(object):
                     server_config_paths)
         self.bclient_config_path_ = client_config_path
 
-    def prepare_server(self, workdir=None, port=9292, device="cpu"):
+    def prepare_server(self,
+                       workdir=None,
+                       port=9292,
+                       device="cpu",
+                       cube_conf=None):
         if not self._port_is_available(port):
             raise SystemExit("Prot {} is already used".format(port))
         default_port = 12000
@@ -656,7 +670,10 @@ class MultiLangServer(object):
                 self.port_list_.append(default_port + i)
                 break
         self.bserver_.prepare_server(
-            workdir=workdir, port=self.port_list_[0], device=device)
+            workdir=workdir,
+            port=self.port_list_[0],
+            device=device,
+            cube_conf=cube_conf)
         self.set_port(port)
 
     def _launch_brpc_service(self, bserver):
diff --git a/python/paddle_serving_server_gpu/__init__.py b/python/paddle_serving_server_gpu/__init__.py
index 26288f6ae65ce823a57ee201130d40ff6510c4a5..df04cb7840bbacd90ccb7e3c66147a6856b23e02 100644
--- a/python/paddle_serving_server_gpu/__init__.py
+++ b/python/paddle_serving_server_gpu/__init__.py
@@ -26,7 +26,7 @@ from contextlib import closing
 import argparse
 import collections
 import fcntl
-
+import shutil
 import numpy as np
 import grpc
 from .proto import multi_lang_general_model_service_pb2
@@ -285,7 +285,7 @@ class Server(object):
             infer_service.workflows.extend(["workflow1"])
             self.infer_service_conf.services.extend([infer_service])
 
-    def _prepare_resource(self, workdir):
+    def _prepare_resource(self, workdir, cube_conf):
         self.workdir = workdir
         if self.resource_conf == None:
             with open("{}/{}".format(workdir, self.general_model_config_fn),
@@ -297,6 +297,11 @@ class Server(object):
                     if "dist_kv" in node.name:
                         self.resource_conf.cube_config_path = workdir
                         self.resource_conf.cube_config_file = self.cube_config_fn
+                        if cube_conf == None:
+                            raise ValueError(
+                                "Please set the path of cube.conf while use dist_kv op."
+                            )
+                        shutil.copy(cube_conf, workdir)
             self.resource_conf.model_toolkit_path = workdir
             self.resource_conf.model_toolkit_file = self.model_toolkit_fn
             self.resource_conf.general_model_path = workdir
@@ -406,7 +411,11 @@ class Server(object):
         os.chdir(self.cur_path)
         self.bin_path = self.server_path + "/serving"
 
-    def prepare_server(self, workdir=None, port=9292, device="cpu"):
+    def prepare_server(self,
+                       workdir=None,
+                       port=9292,
+                       device="cpu",
+                       cube_conf=None):
         if workdir == None:
             workdir = "./tmp"
             os.system("mkdir {}".format(workdir))
@@ -418,7 +427,7 @@ class Server(object):
             raise SystemExit("Port {} is already used".format(port))
 
         self.set_port(port)
-        self._prepare_resource(workdir)
+        self._prepare_resource(workdir, cube_conf)
         self._prepare_engine(self.model_config_paths, device)
         self._prepare_infer_service(port)
         self.workdir = workdir
@@ -690,7 +699,11 @@ class MultiLangServer(object):
                     server_config_paths)
         self.bclient_config_path_ = client_config_path
 
-    def prepare_server(self, workdir=None, port=9292, device="cpu"):
+    def prepare_server(self,
+                       workdir=None,
+                       port=9292,
+                       device="cpu",
+                       cube_conf=None):
         if not self._port_is_available(port):
             raise SystemExit("Prot {} is already used".format(port))
         default_port = 12000
@@ -701,7 +714,10 @@ class MultiLangServer(object):
                 self.port_list_.append(default_port + i)
                 break
         self.bserver_.prepare_server(
-            workdir=workdir, port=self.port_list_[0], device=device)
+            workdir=workdir,
+            port=self.port_list_[0],
+            device=device,
+            cube_conf=cube_conf)
         self.set_port(port)
 
     def _launch_brpc_service(self, bserver):
diff --git a/tools/serving_build.sh b/tools/serving_build.sh
index 097123165988fb266f7c4a3a0da603ade6d98be1..498969b9c886a56e298d397808c2da14a38bac85 100644
--- a/tools/serving_build.sh
+++ b/tools/serving_build.sh
@@ -229,10 +229,7 @@ function python_run_criteo_ctr_with_cube() {
             check_cmd "mv models/data ./cube/"
             check_cmd "mv models/ut_data ./"
             cp ../../../build-server-$TYPE/output/bin/cube* ./cube/
-            mkdir -p $PYTHONROOT/lib/python2.7/site-packages/paddle_serving_server/serving-cpu-avx-openblas-0.1.3/
-            yes | cp ../../../build-server-$TYPE/output/demo/serving/bin/serving $PYTHONROOT/lib/python2.7/site-packages/paddle_serving_server/serving-cpu-avx-openblas-0.1.3/
             sh cube_prepare.sh &
-            check_cmd "mkdir work_dir1 && cp cube/conf/cube.conf ./work_dir1/"
             python test_server.py ctr_serving_model_kv &
             sleep 5
             check_cmd "python test_client.py ctr_client_conf/serving_client_conf.prototxt ./ut_data >score"
@@ -257,10 +254,7 @@ function python_run_criteo_ctr_with_cube() {
             check_cmd "mv models/data ./cube/"
             check_cmd "mv models/ut_data ./"
             cp ../../../build-server-$TYPE/output/bin/cube* ./cube/
-            mkdir -p $PYTHONROOT/lib/python2.7/site-packages/paddle_serving_server_gpu/serving-gpu-0.1.3/
-            yes | cp ../../../build-server-$TYPE/output/demo/serving/bin/serving $PYTHONROOT/lib/python2.7/site-packages/paddle_serving_server_gpu/serving-gpu-0.1.3/
             sh cube_prepare.sh &
-            check_cmd "mkdir work_dir1 && cp cube/conf/cube.conf ./work_dir1/"
             python test_server_gpu.py ctr_serving_model_kv &
             sleep 5
             # for warm up