Merge pull request #6 from PaddlePaddle/master

Sync with remote

doc/ELASTIC_CTR.md

@@ -32,6 +32,8 @@ ELASTIC CTR
 - cube-builder: 负责将训练作业产出的模型文件（hadoop sequence file格式）转换成可以被cube-server加载的字典文件。字典文件具有特定的数据结构，针对尺寸和内存中访问做了高度优化
 - Cube-Server: 提供分片kv读写能力的服务节点
 - Cube-agent: 与cube-server同机部署，接收cube-transfer下发的字典文件更新命令，拉取数据到本地，通知cube-server进行更新
+- Paddle Serving: 加载CTR预估任务模型ProgramDesc和dense参数，提供预测服务
+- Client: CTR预估任务的demo客户端
 
 以上组件串联完成从训练到预测部署的所有流程。本文档所提供的一键部署脚本[paddle-suite.sh](https://github.com/PaddlePaddle/Serving/blob/master/doc/resource/paddle-suite.sh)可一键部署上述所有组件。
 
@@ -121,7 +123,7 @@ $ kubectl apply -f https://raw.githubusercontent.com/volcano-sh/volcano/master/i
 
 ## 3.1 下载部署方案脚本文件
 
-请将[本方案所需所有脚本文件](https://github.com/PaddlePaddle/edl/tree/develop/example/ctr/script)下载到本地
+请将[本方案所需所有脚本文件](https://github.com/PaddlePaddle/Serving/tree/master/doc/resource)下载到本地
 
 ## 3.2 一键部署
 
@@ -133,9 +135,7 @@ $ bash paddle-suite.sh
 
 请参考**3.3-3.8节**验证每一步的安装是否正确，**第4节**验证训练过程和预测服务结果。
 
-任务的所有脚本文件可以访问[这里](https://github.com/PaddlePaddle/edl/tree/develop/example/ctr/script)获取。
-
-**注**：以下**3.3-3.8节所述内容已经在一键部署脚本中包含，无需手动执行**。但为方便理解，将该脚本的每一步执行过程给出说明。
+**[注意！！！]**：以下**3.3-3.8节所述内容已经在一键部署脚本中包含，无需手动执行**。但为方便理解，将该脚本的每一步执行过程给出说明。
 
 ## 3.3 选择一个node作为输出节点
 
@@ -405,16 +405,16 @@ $ ./get_values -h 192.168.1.1 -t 3 -r 10000 -b 1000
 
 并发数 （压测线程数） | batch size | 平均响应时间 (us) | total qps
 -------|------------|-------------|---------------------------
-1  | 1000 | 1159 | 862
-4  | 1000 | 3537  | 1079
-8  | 1000 | 7726  | 1073
-16 | 1000 | 15440  | 1034
-24 | 1000 | 24279  | 1004 
-32 | 1000 | 32570 | 996
+1  | 1000 | 1643 | 608
+4  | 1000 | 4878  | 819
+8  | 1000 | 9870  | 810
+16 | 1000 | 22177  | 721
+24 | 1000 | 30620  | 783 
+32 | 1000 | 37668 | 849
 
 ###测试结论
 
-由于Redis高效的时间驱动模型和全内存操作，在单并发时，redis平均响应时间比cube少接近50% (1100us vs. 1680us)
+由于Redis高效的时间驱动模型和全内存操作，在单并发时，redis平均响应时间与cube相差不多% (1643us vs. 1312us)
 
 在扩展性方面，redis受制于单线程模型，随并发数增加，响应时间加倍增加，而总吞吐在1000qps左右即不再上涨；而cube则随着压测并发数增加，总的qps一直上涨，说明cube能够较好处理并发请求，具有良好的扩展能力。
 

doc/resource/get_value.cpp

@@ -20,6 +20,7 @@ int batch_size = 100;
 int key_size = 10000000;        // keys in redis server
 
 std::vector<uint64_t> times_us;
+std::vector<uint64_t> average_time_us;
 
 sw::redis::Redis *redis;
 
@@ -94,7 +95,7 @@ void thread_worker(int thread_id)
         std::vector<std::string> get_kvs_res;
 
          for(int j = i * batch_size; j <  (i + 1) * batch_size; j++) {
-            get_kvs.push_back(std::to_string(i % key_size));
+            get_kvs.push_back(std::to_string(j % key_size));
         }
         auto start2 = std::chrono::steady_clock::now();
         redis->mget(get_kvs.begin(), get_kvs.end(), std::back_inserter(get_kvs_res));
@@ -102,10 +103,11 @@ void thread_worker(int thread_id)
         times_us[thread_id] += std::chrono::duration_cast<std::chrono::microseconds>(stop2 - start2).count();
     }
 
-    // Per-thread statistics
-    std::cout << total_request_num << " requests, " << batch_size << " keys per req, total time us = " << times_us[thread_id] <<std::endl;
-    std::cout << "Average " << times_us[thread_id] / total_request_num << "us per req" << std::endl;
-    std::cout << "qps: " << (double)total_request_num / times_us[thread_id] * 1000000 << std::endl;
+    average_time_us[thread_id] = times_us[thread_id] / total_request_num;
+
+    // std::cout << total_request_num << " requests, " << batch_size << " keys per req, total time us = " << times_us[thread_id] <<std::endl;
+    // std::cout << "Average " << average_time_us[thread_id] << "us per req" << std::endl;
+    // std::cout << "qps: " << (double)total_request_num / times_us[thread_id] * 1000000 << std::endl;
 }
 
 int main(int argc, char **argv)
@@ -117,6 +119,7 @@ int main(int argc, char **argv)
 
     std::vector<std::thread> workers;
     times_us.reserve(thread_num);
+    average_time_us.reserve(thread_num);
 
     for (int i = 0; i < thread_num; ++i) {
         times_us[i] = 0;
@@ -127,18 +130,19 @@ int main(int argc, char **argv)
         workers[i].join();
     }
 
-    // times_total_us is average running time of each thread
     uint64_t times_total_us = 0;
+    uint64_t average_time_total_us;
+
     for (int i = 0; i < thread_num; ++i) {
         times_total_us += times_us[i];
+        average_time_total_us += average_time_us[i];
     }
-    times_total_us /= thread_num;
 
-    // Total requests should be sum of requests sent by each thread
+    times_total_us /= thread_num;
     total_request_num *= thread_num;
 
     std::cout << total_request_num << " requests, " << batch_size << " keys per req, total time us = " << times_total_us <<std::endl;
-    std::cout << "Average " << times_total_us / total_request_num << "us per req" << std::endl;
+    std::cout << "Average " << average_time_total_us / thread_num << "us per req" << std::endl;
     std::cout << "qps: " << (double)total_request_num / times_total_us * 1000000 << std::endl;
 
     return 0;