Merge pull request #705 from gentelyang/develop

fix benchmark

python/examples/bert/benchmark.py

@@ -116,8 +116,10 @@ def single_func(idx, resource):
 
 if __name__ == '__main__':
     multi_thread_runner = MultiThreadRunner()
-    endpoint_list = ["127.0.0.1:9292"]
-    turns = 10
+    endpoint_list = [
+        "127.0.0.1:9292", "127.0.0.1:9293", "127.0.0.1:9294", "127.0.0.1:9295"
+    ]
+    turns = 100
     start = time.time()
     result = multi_thread_runner.run(
         single_func, args.thread, {"endpoint": endpoint_list,
@@ -130,9 +132,9 @@ if __name__ == '__main__':
         avg_cost += result[0][i]
     avg_cost = avg_cost / args.thread
 
-    print("total cost :{} s".format(total_cost))
-    print("each thread cost :{} s. ".format(avg_cost))
-    print("qps :{} samples/s".format(args.batch_size * args.thread * turns /
-                                     total_cost))
+    print("total cost: {}s".format(total_cost))
+    print("each thread cost: {}s. ".format(avg_cost))
+    print("qps: {}samples/s".format(args.batch_size * args.thread * turns /
+                                    total_cost))
     if os.getenv("FLAGS_serving_latency"):
         show_latency(result[1])

python/examples/bert/benchmark.sh

@@ -4,8 +4,9 @@ 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
 
 #warm up
 python3 benchmark.py --thread 8 --batch_size 1 --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
@@ -14,14 +15,24 @@ for thread_num in 4 8 16
 do
 for batch_size in 1 4 16 64 256
 do
+    job_bt=`date '+%Y%m%d%H%M%S'`
+    nvidia-smi --id=$gpu_id --query-compute-apps=used_memory --format=csv -lms 100 > gpu_use.log 2>&1 &
+    gpu_memory_pid=$!
     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
+    kill ${gpu_memory_pid}
+    echo "model_name:" $1
+    echo "thread_num:" $thread_num
+    echo "batch_size:" $batch_size
     echo "=================Done===================="
-    echo "model name :$1" >> profile_log_$1
-    echo "batch size :$batch_size" >> profile_log_$1
-    python3 ../util/show_profile.py profile $thread_num >> profile_log_$1
+    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
     tail -n 8 profile >> profile_log_$1
     echo "" >> profile_log_$1
 done

python/examples/imdb/benchmark.py

@@ -13,13 +13,14 @@
 # limitations under the License.
 # pylint: disable=doc-string-missing
 
+import os
 import sys
 import time
 import requests
 from paddle_serving_app.reader import IMDBDataset
 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 MultiThreadRunner, benchmark_args, show_latency
 
 args = benchmark_args()
 
@@ -31,6 +32,13 @@ def single_func(idx, resource):
     with open("./test_data/part-0") as fin:
         for line in fin:
             dataset.append(line.strip())
+    profile_flags = False
+    latency_flags = False
+    if os.getenv("FLAGS_profile_client"):
+        profile_flags = True
+    if os.getenv("FLAGS_serving_latency"):
+        latency_flags = True
+        latency_list = []
     start = time.time()
     if args.request == "rpc":
         client = Client()
@@ -67,9 +75,26 @@ def single_func(idx, resource):
     return [[end - start]]
 
 
-multi_thread_runner = MultiThreadRunner()
-result = multi_thread_runner.run(single_func, args.thread, {})
-avg_cost = 0
-for cost in result[0]:
-    avg_cost += cost
-print("total cost {} s of each thread".format(avg_cost / args.thread))
+if __name__ == '__main__':
+    multi_thread_runner = MultiThreadRunner()
+    endpoint_list = [
+        "127.0.0.1:9292", "127.0.0.1:9293", "127.0.0.1:9294", "127.0.0.1:9295"
+    ]
+    turns = 100
+    start = time.time()
+    result = multi_thread_runner.run(
+        single_func, args.thread, {"endpoint": endpoint_list,
+                                   "turns": turns})
+    end = time.time()
+    total_cost = end - start
+    avg_cost = 0
+    for i in range(args.thread):
+        avg_cost += result[0][i]
+    avg_cost = avg_cost / args.thread
+
+    print("total cost: {}".format(total_cost))
+    print("each thread cost: {}".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[0])

python/examples/imdb/benchmark.sh

 rm profile_log
-for thread_num in 1 2 4 8 16
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+export FLAGS_profile_server=1
+export FLAGS_profile_client=1
+export FLAGS_serving_latency=1
+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 &
+hostname=`echo $(hostname)|awk -F '.baidu.com' '{print $1}'`
+
+sleep 5
+
+for thread_num in 4 8 16
 do
-for batch_size in 1 2 4 8 16 32 64 128 256 512
+for batch_size in 1 4 16 64 256
 do
-    $PYTHONROOT/bin/python benchmark.py --thread $thread_num --batch_size $batch_size --model imdb_bow_client_conf/serving_client_conf.prototxt --request rpc > profile 2>&1
-    echo "========================================"
-    echo "batch size : $batch_size" >> profile_log
-    $PYTHONROOT/bin/python ../util/show_profile.py profile $thread_num >> profile_log
-    tail -n 1 profile >> profile_log
+    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
+    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
+    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
+    tail -n 8 profile >> profile_log_$1
+    echo "" >> profile_log_$1
 done
 done
+
+ps -ef|grep 'serving'|grep -v grep|cut -c 9-15 | xargs kill -9

python/examples/util/show_profile.py

@@ -31,7 +31,7 @@ with open(profile_file) as f:
         if line[0] == "PROFILE":
             prase(line[2])
 
-print("thread num :{}".format(thread_num))
+print("thread_num: {}".format(thread_num))
 for name in time_dict:
-    print("{} cost :{} s in each thread ".format(name, time_dict[name] / (
+    print("{} cost: {}s in each thread ".format(name, time_dict[name] / (
         1000000.0 * float(thread_num))))

python/paddle_serving_client/utils/__init__.py

@@ -39,11 +39,11 @@ def benchmark_args():
 def show_latency(latency_list):
     latency_array = np.array(latency_list)
     info = "latency:\n"
-    info += "mean :{} ms\n".format(np.mean(latency_array))
-    info += "median :{} ms\n".format(np.median(latency_array))
-    info += "80 percent :{} ms\n".format(np.percentile(latency_array, 80))
-    info += "90 percent :{} ms\n".format(np.percentile(latency_array, 90))
-    info += "99 percent :{} ms\n".format(np.percentile(latency_array, 99))
+    info += "mean: {}ms\n".format(np.mean(latency_array))
+    info += "median: {}ms\n".format(np.median(latency_array))
+    info += "80 percent: {}ms\n".format(np.percentile(latency_array, 80))
+    info += "90 percent: {}ms\n".format(np.percentile(latency_array, 90))
+    info += "99 percent: {}ms\n".format(np.percentile(latency_array, 99))
     sys.stderr.write(info)