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。。。
Official code of paper "PGT: A Progressive Method for Training Models on Long Videos" on CVPR2021

PGT Code for paper PGT: A Progressive Method for Training Models on Long Videos. Install Run pip install -r requirements.txt. Run python setup.py buil

Bo Pang 27 Mar 30, 2022
Multispectral Object Detection with Yolov5

Multispectral-Object-Detection Intro Official Code for Cross-Modality Fusion Transformer for Multispectral Object Detection. Multispectral Object Dete

Richard Fang 121 Jan 01, 2023
This repo provides the official code for TransBTS: Multimodal Brain Tumor Segmentation Using Transformer (https://arxiv.org/pdf/2103.04430.pdf).

TransBTS: Multimodal Brain Tumor Segmentation Using Transformer This repo is the official implementation for TransBTS: Multimodal Brain Tumor Segmenta

Raymond 247 Dec 28, 2022
Using VideoBERT to tackle video prediction

VideoBERT This repo reproduces the results of VideoBERT (https://arxiv.org/pdf/1904.01766.pdf). Inspiration was taken from https://github.com/MDSKUL/M

75 Dec 14, 2022
OpenVINO黑客松比赛项目

Window_Guard OpenVINO黑客松比赛项目 英文名称:Window_Guard 中文名称:窗口卫士 硬件 树莓派4B 8G版本 一个磁石开关 USB摄像头(MP4视频文件也可以) 软件(库) OpenVINO RPi 使用方法 本项目使用的OPenVINO是是2021.3版本,并使用了

Tango 6 Jul 04, 2021
Self-Learning - Books Papers, Courses & more I have to learn soon

Self-Learning This repository is intended to be used for personal use, all rights reserved to respective owners, please cite original authors and ask

Achint Chaudhary 968 Jan 02, 2022
Voice Conversion Using Speech-to-Speech Neuro-Style Transfer

This repo contains the official implementation of the VAE-GAN from the INTERSPEECH 2020 paper Voice Conversion Using Speech-to-Speech Neuro-Style Transfer.

Ehab AlBadawy 93 Jan 05, 2023
[제 13회 투빅스 컨퍼런스] OK Mugle! - 장르부터 멜로디까지, Content-based Music Recommendation

Ok Mugle! 🎵 장르부터 멜로디까지, Content-based Music Recommendation 'Ok Mugle!'은 제13회 투빅스 컨퍼런스(2022.01.15)에서 진행한 음악 추천 프로젝트입니다. Description 📖 본 프로젝트에서는 Kakao

SeongBeomLEE 5 Oct 09, 2022
Live training loss plot in Jupyter Notebook for Keras, PyTorch and others

livelossplot Don't train deep learning models blindfolded! Be impatient and look at each epoch of your training! (RECENT CHANGES, EXAMPLES IN COLAB, A

Piotr Migdał 1.2k Jan 08, 2023
GANmouflage: 3D Object Nondetection with Texture Fields

GANmouflage: 3D Object Nondetection with Texture Fields Rui Guo1 Jasmine Collins

29 Aug 10, 2022
An extremely simple, intuitive, hardware-friendly, and well-performing network structure for LiDAR semantic segmentation on 2D range image. IROS21

FIDNet_SemanticKITTI Motivation Implementing complicated network modules with only one or two points improvement on hardware is tedious. So here we pr

YimingZhao 54 Dec 12, 2022
Codes for the compilation and visualization examples to the HIF vegetation dataset

High-impedance vegetation fault dataset This repository contains the codes that compile the "Vegetation Conduction Ignition Test Report" data, which a

1 Dec 12, 2021
Java and SHACL code commented in the paper "Towards compliance checking in reified I/O logic via SHACL" submitted to ICAIL 2021

shRIOL The subfolder shRIOL contains Java files to execute the SHACL files on the OWL ontology. To compile the Java files: "javac -cp ./src/;./lib/* -

1 Dec 06, 2022
Deep Learning Head Pose Estimation using PyTorch.

Hopenet is an accurate and easy to use head pose estimation network. Models have been trained on the 300W-LP dataset and have been tested on real data with good qualitative performance.

Nataniel Ruiz 1.3k Dec 26, 2022
A PyTorch Implementation of Gated Graph Sequence Neural Networks (GGNN)

A PyTorch Implementation of GGNN This is a PyTorch implementation of the Gated Graph Sequence Neural Networks (GGNN) as described in the paper Gated G

Ching-Yao Chuang 427 Dec 13, 2022
This repository is an implementation of paper : Improving the Training of Graph Neural Networks with Consistency Regularization

CRGNN Paper : Improving the Training of Graph Neural Networks with Consistency Regularization Environments Implementing environment: GeForce RTX™ 3090

THUDM 28 Dec 09, 2022
a short visualisation script for pyvideo data

PyVideo Speakers A CLI that visualises repeat speakers from events listed in https://github.com/pyvideo/data Not terribly efficient, but you know. Ins

Katie McLaughlin 3 Nov 24, 2021
This repo will contain code to reproduce and build upon understanding transfer learning

What is being transferred in transfer learning? This repo contains the code for the following paper: Behnam Neyshabur*, Hanie Sedghi*, Chiyuan Zhang*.

4 Jun 16, 2021
deep-table implements various state-of-the-art deep learning and self-supervised learning algorithms for tabular data using PyTorch.

deep-table implements various state-of-the-art deep learning and self-supervised learning algorithms for tabular data using PyTorch.

63 Oct 17, 2022
Research code for Arxiv paper "Camera Motion Agnostic 3D Human Pose Estimation"

GMR(Camera Motion Agnostic 3D Human Pose Estimation) This repo provides the source code of our arXiv paper: Seong Hyun Kim, Sunwon Jeong, Sungbum Park

Seong Hyun Kim 1 Feb 07, 2022