Simple Dynamic Batching Inference

Related tags

Deep LearningSDBI
Overview

Simple Dynamic Batching Inference

解决了什么问题?

众所周知,Batch对于GPU上深度学习模型的运行效率影响很大。。。

是在Inference时。搜索、推荐等场景自带比较大的batch,问题不大。但更多场景面临的往往是稀碎的请求(比如图片服务里一次一张图)。

如果想提高服务的吞吐,把稀碎的请求动态攒成Batch再送GPU处理就是刚需。

NV的Triton包含了Dynamic Batching功能。我也用cpp写过一版。但是发现在部署、特别是给别人用python来调用的时候,始终是比较麻烦的。比如要各种配置环境或用NGC的镜像、走个本地rpc等。。

反过来想,只要程序瓶颈还卡在计算上,就有机会用python写一版至少吞吐上可以打平cpp的Dynamic Batching。好处是使用会方便很多。

出于个人需要和兴趣,之前基于multiprocess.Queue写过一版Dynamic Batching。但是Queue本身对于延迟的影响非常大,数字比较难看。

最近发现Python 3.8支持了共享内存,用python写了个基于SharedMemory的Dynamic Batching。

跟大家分享一下效果。

测试环境

模型Resnet50,输入(N,3,224,224)。使用某云的V100。

测试结果

我们先测一下Torch性能上限,好对数据有个基本了解。

然后一步步看不同功能的影响。

对应测试命令:

# 生成一个假模型
python fake_resnet50.py
# 测试
python benchmark.py  --no_dynamic_batch --worker_num=N --worker_batch=M

MPS

多进程Torch + MPS。

进程数量 Batch Latency Throughput
1 1 4.54 ms 220.10 pic/s
4 1 8.05 ms 496.52 pic/s
8 1 13.97 ms 572.57 pic/s
16 1 28.15 ms 526.42 pic/s

可以看出MPS是很有效的,没有MPS时,多进程轮占时间片,多个进程吞吐基本也就卡在200多。

加了多进程后,多进程的kernel在同一context下调度。在8的时候达到最高。

Batching

基于以上数据,再看下Batching的影响。

进程数量 Batch Latency Throughput
4 1 8.05 ms 496.52 pic/s
1 4 6.43 ms 622.07 pic/s
进程数量 Batch Latency Throughput
8 1 13.97 ms 572.57 pic/s
1 8 10.43 ms 766.93 pic/s
进程数量 Batch Latency Throughput
16 1 28.15 ms 526.42 pic/s
1 16 18.03 ms 887.20 pic/s

可以看到MPS虽然对吞吐有帮助,但是有条件的话,Batching依旧是更好的选择。

MPS+Batching测Torch上限

在测一下Batch=32(或者其他比较高的数字都可),看一下torch框架的上限。

进程数量 Batch Latency Throughput
1 32 33.54 ms 953.60 pic/s
2 32 56.98 ms 1123.20 pic/s
3 32 78.96 ms 1215.47 pic/s
4 32 109.89 ms 1164.80 pic/s

即便batch比较大了,但MPS依旧有提升。

Dynamic Batching

实际应用中,琐碎请求会带来的性能下降。如果对于延迟的要求没有非常苛刻,那么是可以通过牺牲一部分延迟(用来打Batch),换取更高的吞吐(省钱)。

所以这轮测试的场景是,有N个数据(业务)进程,每个进程数据batch=1,达到MPS+Batching的上限吞吐。

先试一下对上述最大吞吐的case。128个数据(业务)进程,每个进程灌一张图,后台通过共享内存传输数据并打batch。

测试命令:

python benchmark.py --worker_num=128 --worker_batch=1 --max_batch_size=32 --model_num=3 --wait_time=0.01
数据(业务)进程 GPU模型进程 Latency Throughput
128 3 103.45 ms 1237.33 pic/s

能够达到极限延迟,但比最理想的情况增加了20%+的延迟。

找个小的场景试一下:

python benchmark.py --worker_num=8 --worker_batch=1 --max_batch_size=4 --model_num=2 --wait_time=0.003
数据(业务)进程 GPU模型进程 Latency Throughput
8 2 13.04 ms 613.40 pic/s

跟前面Torch测试的数字对比,可以理解成这case下8个请求进程被分成两组,总体基本能够达到batch=4的吞吐。

时间都去哪了?

针对1200+的最大吞吐场景分析了一下:

延迟由 batch + MPS 的 79 ms 增加至 Dynamic Batching 的 103ms.其中,

  • 19ms 左右是拼batch的时间,其中10ms是命令中的等待时间,还有8.3ms的np.concat时间。
  • 分割输出回各数据进程大概用了1ms。
  • 各种队列的等待时间。

总的来说没有不太合理的地方,在benchmark里我也把各部分时间收集和打出来了。

施工图

施工图

虽然源码不长(<1000行),结构也简单。但各种进程和通信还是有点多的。

程序启动时创建context进程,每个数据进程创建模型实例时:

  • context 进程会查看是否已存在对应的模型backend进程
    • 存在 -> 通过shared memory 建立连接
    • 不存在 -> 创建backend进程 -> 创建模型进程
  • 多个模型进程是为了充分利用MPS
  • 当用户进程中有多段模型时,会创建相应多个backend进程,比如识别+检测等等
  • 进程间不传输数据,仅传输shared memory地址和tensor元信息。

代码 & 相关说明

原理大概就是这个 shared_memory sample

测试代码:benchmark.py

使用样例:sample.py

  • 基本跟用pytorch差不多,load+forward。但是:
    • 要指定数据最大尺寸,用来分配shared memory
    • 最后要用一个Run函数启动,因为要提前初始化一些进程变量
    • 需要为模型指定name。当程序涉及到多个模型的时候,数据进程通过name连接到特定的模型进程。

Konwn issues

multiprocess.shared_memory在回收时,在一些系统下会报leak或已经释放的error/warning,一些系统正常。

错的系统我跑官方示例也有错。所以还不好判断是什么原因。如果觉得可以忍又不想烦可以用下面的命令禁掉。

export PYTHONWARNINGS=ignore

最后

If 有人感兴趣 and 我有时间

  • 支持一下TensorRT/TensorCore FP16,以及某个特定版本的TF。
  • 输出还没有全用shared memory(主要是我懒),所以大输出模型的 吞吐/延迟 会受到数据拷贝的影响。可以改进。。。
For the paper entitled ''A Case Study and Qualitative Analysis of Simple Cross-Lingual Opinion Mining''

Summary This is the source code for the paper "A Case Study and Qualitative Analysis of Simple Cross-Lingual Opinion Mining", which was accepted as fu

1 Nov 10, 2021
Dataset Condensation with Contrastive Signals

Dataset Condensation with Contrastive Signals This repository is the official implementation of Dataset Condensation with Contrastive Signals (DCC). T

3 May 19, 2022
Official implementation for: Blended Diffusion for Text-driven Editing of Natural Images.

Blended Diffusion for Text-driven Editing of Natural Images Blended Diffusion for Text-driven Editing of Natural Images Omri Avrahami, Dani Lischinski

328 Dec 30, 2022
Research Artifact of USENIX Security 2022 Paper: Automated Side Channel Analysis of Media Software with Manifold Learning

Automated Side Channel Analysis of Media Software with Manifold Learning Official implementation of USENIX Security 2022 paper: Automated Side Channel

Yuanyuan Yuan 175 Jan 07, 2023
Official implementation of "SinIR: Efficient General Image Manipulation with Single Image Reconstruction" (ICML 2021)

SinIR (Official Implementation) Requirements To install requirements: pip install -r requirements.txt We used Python 3.7.4 and f-strings which are in

47 Oct 11, 2022
PyTorch GPU implementation of the ES-RNN model for time series forecasting

Fast ES-RNN: A GPU Implementation of the ES-RNN Algorithm A GPU-enabled version of the hybrid ES-RNN model by Slawek et al that won the M4 time-series

Kaung 305 Jan 03, 2023
Collection of sports betting AI tools.

sports-betting sports-betting is a collection of tools that makes it easy to create machine learning models for sports betting and evaluate their perf

George Douzas 109 Dec 31, 2022
CCP dataset from Clothing Co-Parsing by Joint Image Segmentation and Labeling

Clothing Co-Parsing (CCP) Dataset Clothing Co-Parsing (CCP) dataset is a new clothing database including elaborately annotated clothing items. 2, 098

Wei Yang 434 Dec 24, 2022
Deep Learning and Reinforcement Learning Library for Scientists and Engineers 🔥

TensorLayer is a novel TensorFlow-based deep learning and reinforcement learning library designed for researchers and engineers. It provides an extens

TensorLayer Community 7.1k Dec 29, 2022
Ἀνατομή is a PyTorch library to analyze representation of neural networks

Ἀνατομή is a PyTorch library to analyze representation of neural networks

Ryuichiro Hataya 50 Dec 05, 2022
DatasetGAN: Efficient Labeled Data Factory with Minimal Human Effort

DatasetGAN This is the official code and data release for: DatasetGAN: Efficient Labeled Data Factory with Minimal Human Effort Yuxuan Zhang*, Huan Li

302 Jan 05, 2023
Code for the paper 'A High Performance CRF Model for Clothes Parsing'.

Clothes Parsing Overview This code provides an implementation of the research paper: A High Performance CRF Model for Clothes Parsing Edgar Simo-S

Edgar Simo-Serra 119 Nov 21, 2022
AdamW optimizer for bfloat16 models in pytorch.

Image source AdamW optimizer for bfloat16 models in pytorch. Bfloat16 is currently an optimal tradeoff between range and relative error for deep netwo

Alex Rogozhnikov 8 Nov 20, 2022
A basic reminder tool written in Python.

A simple Python Reminder Here's a basic reminder tool written in Python that speaks to the user and sends a notification. Run pip3 install pyttsx3 w

Sachit Yadav 4 Feb 05, 2022
Algorithmic Trading using RNN

Deep-Trading This an implementation adapted from Rachnog Neural networks for algorithmic trading. Part One — Simple time series forecasting and this c

Hazem Nomer 29 Sep 04, 2022
SciPy fixes and extensions

scipyx SciPy is large library used everywhere in scientific computing. That's why breaking backwards-compatibility comes as a significant cost and is

Nico Schlömer 16 Jul 17, 2022
Evaluating Cross-lingual Sentence Representations

XNLI: The Cross-Lingual NLI Corpus XNLI is an evaluation corpus for language transfer and cross-lingual sentence classification in 15 languages. New:

Meta Research 395 Dec 19, 2022
Uncertain natural language inference

Uncertain Natural Language Inference This repository hosts the code for the following paper: Tongfei Chen*, Zhengping Jiang*, Adam Poliak, Keisuke Sak

Tongfei Chen 14 Sep 01, 2022
DeepMoCap: Deep Optical Motion Capture using multiple Depth Sensors and Retro-reflectors

DeepMoCap: Deep Optical Motion Capture using multiple Depth Sensors and Retro-reflectors By Anargyros Chatzitofis, Dimitris Zarpalas, Stefanos Kollias

tofis 24 Oct 08, 2022
Reinforcement learning library(framework) designed for PyTorch, implements DQN, DDPG, A2C, PPO, SAC, MADDPG, A3C, APEX, IMPALA ...

Automatic, Readable, Reusable, Extendable Machin is a reinforcement library designed for pytorch. Build status Platform Status Linux Windows Supported

Iffi 348 Dec 24, 2022