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。。。
This a classic fintech problem that introduces real life difficulties such as data imbalance. Check out the notebook to find out more!

Credit Card Fraud Detection Introduction Online transactions have become a crucial part of any business over the years. Many of those transactions use

Jonathan Hasbani 0 Jan 20, 2022
[NeurIPS 2021 Spotlight] Aligning Pretraining for Detection via Object-Level Contrastive Learning

SoCo [NeurIPS 2021 Spotlight] Aligning Pretraining for Detection via Object-Level Contrastive Learning By Fangyun Wei*, Yue Gao*, Zhirong Wu, Han Hu,

Yue Gao 139 Dec 14, 2022
Nonuniform-to-Uniform Quantization: Towards Accurate Quantization via Generalized Straight-Through Estimation. In CVPR 2022.

Nonuniform-to-Uniform Quantization This repository contains the training code of N2UQ introduced in our CVPR 2022 paper: "Nonuniform-to-Uniform Quanti

Zechun Liu 60 Dec 28, 2022
Python KNN model: Predicting a probability of getting a work visa. Tableau: Non-immigrant visas over the years.

The value of international students to the United States. Probability of getting a non-immigrant visa. Project timeline: Jan 2021 - April 2021 Project

Zinaida Dvoskina 2 Nov 21, 2021
GANSketchingJittor - Implementation of Sketch Your Own GAN in Jittor

GANSketching in Jittor Implementation of (Sketch Your Own GAN) in Jittor(计图). Or

Bernard Tan 10 Jul 02, 2022
This repository contains the code for the ICCV 2019 paper "Occupancy Flow - 4D Reconstruction by Learning Particle Dynamics"

Occupancy Flow This repository contains the code for the project Occupancy Flow - 4D Reconstruction by Learning Particle Dynamics. You can find detail

189 Dec 29, 2022
An Industrial Grade Federated Learning Framework

DOC | Quick Start | 中文 FATE (Federated AI Technology Enabler) is an open-source project initiated by Webank's AI Department to provide a secure comput

Federated AI Ecosystem 4.8k Jan 09, 2023
🚀 PyTorch Implementation of "Progressive Distillation for Fast Sampling of Diffusion Models(v-diffusion)"

PyTorch Implementation of "Progressive Distillation for Fast Sampling of Diffusion Models(v-diffusion)" Unofficial PyTorch Implementation of Progressi

Vitaliy Hramchenko 58 Dec 19, 2022
This repository contains small projects related to Neural Networks and Deep Learning in general.

ILearnDeepLearning.py Description People say that nothing develops and teaches you like getting your hands dirty. This repository contains small proje

Piotr Skalski 1.2k Dec 22, 2022
Unsupervised Image to Image Translation with Generative Adversarial Networks

Unsupervised Image to Image Translation with Generative Adversarial Networks Paper: Unsupervised Image to Image Translation with Generative Adversaria

Hao 71 Oct 30, 2022
Time Dependent DFT in Tamm-Dancoff Approximation

Density Function Theory Program - kspy-tddft(tda) This is an implementation of Time-Dependent Density Functional Theory(TDDFT) using the Tamm-Dancoff

Peter Borthwick 2 Nov 17, 2022
Official Repsoitory for "Mish: A Self Regularized Non-Monotonic Neural Activation Function" [BMVC 2020]

Mish: Self Regularized Non-Monotonic Activation Function BMVC 2020 (Official Paper) Notes: (Click to expand) A considerably faster version based on CU

Xa9aX ツ 1.2k Dec 29, 2022
Group-Free 3D Object Detection via Transformers

Group-Free 3D Object Detection via Transformers By Ze Liu, Zheng Zhang, Yue Cao, Han Hu, Xin Tong. This repo is the official implementation of "Group-

Ze Liu 213 Dec 07, 2022
EquiBind: Geometric Deep Learning for Drug Binding Structure Prediction

EquiBind: geometric deep learning for fast predictions of the 3D structure in which a small molecule binds to a protein

Hannes Stärk 355 Jan 03, 2023
Unofficial Implementation of RobustSTL: A Robust Seasonal-Trend Decomposition Algorithm for Long Time Series (AAAI 2019)

RobustSTL: A Robust Seasonal-Trend Decomposition Algorithm for Long Time Series (AAAI 2019) This repository contains python (3.5.2) implementation of

Doyup Lee 222 Dec 21, 2022
A Next Generation ConvNet by FaceBookResearch Implementation in PyTorch(Original) and TensorFlow.

ConvNeXt A Next Generation ConvNet by FaceBookResearch Implementation in PyTorch(Original) and TensorFlow. A FacebookResearch Implementation on A Conv

Raghvender 2 Feb 14, 2022
A solution to the 2D Ising model of ferromagnetism, implemented using the Metropolis algorithm

Solving the Ising model on a 2D lattice using the Metropolis Algorithm Introduction The Ising model is a simplified model of ferromagnetism, the pheno

Rohit Prabhu 5 Nov 13, 2022
An implementation of the 1. Parallel, 2. Streaming, 3. Randomized SVD using MPI4Py

PYPARSVD This implementation allows for a singular value decomposition which is: Distributed using MPI4Py Streaming - data can be shown in batches to

Romit Maulik 44 Dec 31, 2022
PyTorch Connectomics: segmentation toolbox for EM connectomics

Introduction The field of connectomics aims to reconstruct the wiring diagram of the brain by mapping the neural connections at the level of individua

Zudi Lin 132 Dec 26, 2022
A scikit-learn compatible neural network library that wraps PyTorch

A scikit-learn compatible neural network library that wraps PyTorch. Resources Documentation Source Code Examples To see more elaborate examples, look

4.9k Jan 03, 2023