fix paddle_serving_server.serve

if a user does not define gpus_ids, current api will find gpus through CUDA_VISIBLE_DEVICES
otherwise, the server will startup cpu services as needed.

README.md

@@ -93,6 +93,8 @@ print(fetch_map)
 
 [Compile from source code(Chinese)](doc/COMPILE.md)
 
+[How profile serving efficiency?(Chinese)](https://github.com/PaddlePaddle/Serving/tree/develop/python/examples/util)
+
 [FAQ(Chinese)](doc/FAQ.md)
 
 [Design Doc(Chinese)](doc/DESIGN.md)

python/examples/bert/benchmark.py

@@ -34,7 +34,7 @@ args = benchmark_args()
 def single_func(idx, resource):
     fin = open("data-c.txt")
     if args.request == "rpc":
-        reader = BertReader(vocab_file="vocab.txt", max_seq_len=20)
+        reader = BertReader(vocab_file="vocab.txt", max_seq_len=128)
         config_file = './serving_client_conf/serving_client_conf.prototxt'
         fetch = ["pooled_output"]
         client = Client()
@@ -50,7 +50,6 @@ def single_func(idx, resource):
         start = time.time()
         header = {"Content-Type": "application/json"}
         for line in fin:
-            #dict_data = {"words": "this is for output ", "fetch": ["pooled_output"]}
             dict_data = {"words": line, "fetch": ["pooled_output"]}
             r = requests.post(
                 'http://{}/bert/prediction'.format(resource["endpoint"][0]),
@@ -62,10 +61,11 @@ def single_func(idx, resource):
 
 if __name__ == '__main__':
     multi_thread_runner = MultiThreadRunner()
-    endpoint_list = [
-        "127.0.0.1:9494", "127.0.0.1:9495", "127.0.0.1:9496", "127.0.0.1:9497"
-    ]
-    #endpoint_list = endpoint_list + endpoint_list + endpoint_list
-    #result = multi_thread_runner.run(single_func, args.thread, {"endpoint":endpoint_list})
-    result = single_func(0, {"endpoint": endpoint_list})
+    endpoint_list = []
+    card_num = 4
+    for i in range(args.thread):
+        endpoint_list.append("127.0.0.1:{}".format(9494 + i % card_num))
+    print(endpoint_list)
+    result = multi_thread_runner.run(single_func, args.thread,
+                                     {"endpoint": endpoint_list})
     print(result)

python/paddle_serving_server_gpu/serve.py

@@ -17,8 +17,8 @@ Usage:
     Example:
         python -m paddle_serving_server.serve --model ./serving_server_model --port 9292
 """
-import os
 import argparse
+import os
 from multiprocessing import Pool, Process
 from paddle_serving_server_gpu import serve_args
 
@@ -64,12 +64,14 @@ def start_gpu_card_model(gpuid, args):  # pylint: disable=doc-string-missing
 def start_multi_card(args):  # pylint: disable=doc-string-missing
     gpus = ""
     if args.gpu_ids == "":
-        import os
+        if "CUDA_VISIBLE_DEVICES" in os.environ:
             gpus = os.environ["CUDA_VISIBLE_DEVICES"]
+        else:
+            gpus = []
     else:
         gpus = args.gpu_ids.split(",")
     if len(gpus) <= 0:
-        start_gpu_card_model(-1)
+        start_gpu_card_model(-1, args)
     else:
         gpu_processes = []
         for i, gpu_id in enumerate(gpus):