@@ -37,14 +37,14 @@ We consider deploying deep learning inference service online to be a user-facing
We **highly recommend** you to **run Paddle Serving in Docker**, please visit [Run in Docker](https://github.com/PaddlePaddle/Serving/blob/develop/doc/RUN_IN_DOCKER.md)
This document will take Python2 as an example to show how to compile Paddle Serving. If you want to compile with Python3, just adjust the Python options of cmake:
Due to different model structures, different prediction services consume different computing resources when performing predictions. For online prediction services, models that require less computing resources will have a higher proportion of communication time cost, which is called communication-intensive service. Models that require more computing resources have a higher time cost for inference calculations, which is called computationa-intensive services.
For a prediction service, the easiest way to determine what type it is is to look at the time ratio. Paddle Serving provides [Timeline tool](../python/examples/util/README_CN.md), which can intuitively display the time spent in each stage of the prediction service.
For a prediction service, the easiest way to determine what type it is is to look at the time ratio. Paddle Serving provides [Timeline tool](../python/examples/util/README_CN.md), which can intuitively display the time spent in each stage of the prediction service.
For communication-intensive prediction services, requests can be aggregated, and within a limit that can tolerate delay, multiple prediction requests can be combined into a batch for prediction.
@@ -71,28 +71,3 @@ set environmental variable to specify which gpus are used, the command above mea
```
curl -H "Content-Type:application/json" -X POST -d '{"feed":[{"words": "hello"}], "fetch":["pooled_output"]}' http://127.0.0.1:9292/bert/prediction
```
### Benchmark
Model：bert_chinese_L-12_H-768_A-12
GPU：GPU V100 * 1
CUDA/cudnn Version：CUDA 9.2，cudnn 7.1.4
In the test, 10 thousand samples in the sample data are copied into 100 thousand samples. Each client thread sends a sample of the number of threads. The batch size is 1, the max_seq_len is 20(not 128 as described above), and the time unit is seconds.
When the number of client threads is 4, the prediction speed can reach 432 samples per second.
Because a single GPU can only perform serial calculations internally, increasing the number of client threads can only reduce the idle time of the GPU. Therefore, after the number of threads reaches 4, the increase in the number of threads does not improve the prediction speed.
| client thread num | prepro | client infer | op0 | op1 | op2 | postpro | total |
curl -H "Content-Type:application/json" -X POST -d '{"feed":[{"words": "i am very sad | 0"}], "fetch":["prediction"]}' http://127.0.0.1:9292/imdb/prediction
```
### Benchmark
CPU ：Intel(R) Xeon(R) Gold 6271 CPU @ 2.60GHz * 48
Model ：[CNN](https://github.com/PaddlePaddle/Serving/blob/develop/python/examples/imdb/nets.py)
server thread num ： 16
In this test, client sends 25000 test samples totally, the bar chart given later is the latency of single thread, the unit is second, from which we know the predict efficiency is improved greatly by multi-thread compared to single-thread. 8.7 times improvement is made by 16 threads prediction.
| client thread num | prepro | client infer | op0 | op1 | op2 | postpro | total |
curl -H "Content-Type:application/json" -X POST -d '{"feed":[{"words": "i am very sad | 0"}], "fetch":["prediction"]}' http://127.0.0.1:9292/imdb/prediction
