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。。。
Predicting the duration of arrival delays for commercial flights.

Flight Delay Prediction Our objective is to predict arrival delays of commercial flights. According to the US Department of Transportation, about 21%

Jordan Silke 1 Jan 11, 2022
High-resolution networks and Segmentation Transformer for Semantic Segmentation

High-resolution networks and Segmentation Transformer for Semantic Segmentation Branches This is the implementation for HRNet + OCR. The PyTroch 1.1 v

HRNet 2.8k Jan 07, 2023
The Pytorch implementation for "Video-Text Pre-training with Learned Regions"

Region_Learner The Pytorch implementation for "Video-Text Pre-training with Learned Regions" (arxiv) We are still cleaning up the code further and pre

Rui Yan 0 Mar 20, 2022
Py-FEAT: Python Facial Expression Analysis Toolbox

Py-FEAT is a suite for facial expressions (FEX) research written in Python. This package includes tools to detect faces, extract emotional facial expressions (e.g., happiness, sadness, anger), facial

Computational Social Affective Neuroscience Laboratory 147 Jan 06, 2023
Animation of solving the traveling salesman problem to optimality using mixed-integer programming and iteratively eliminating sub tours

tsp-streamlit Animation of solving the traveling salesman problem to optimality using mixed-integer programming and iteratively eliminating sub tours.

4 Nov 05, 2022
Reimplement of SimSwap training code

SimSwap-train Reimplement of SimSwap training code Instructions 1.Environment Preparation (1)Refer to the README document of SIMSWAP to configure the

seeprettyface.com 111 Dec 31, 2022
An end-to-end library for editing and rendering motion of 3D characters with deep learning [SIGGRAPH 2020]

Deep-motion-editing This library provides fundamental and advanced functions to work with 3D character animation in deep learning with Pytorch. The co

1.2k Dec 29, 2022
Pytorch based library to rank predicted bounding boxes using text/image user's prompts.

pytorch_clip_bbox: Implementation of the CLIP guided bbox ranking for Object Detection. Pytorch based library to rank predicted bounding boxes using t

Sergei Belousov 50 Nov 27, 2022
To propose and implement a multi-class classification approach to disaster assessment from the given data set of post-earthquake satellite imagery.

To propose and implement a multi-class classification approach to disaster assessment from the given data set of post-earthquake satellite imagery.

Kunal Wadhwa 2 Jan 05, 2022
The code for 'Deep Residual Fourier Transformation for Single Image Deblurring'

Deep Residual Fourier Transformation for Single Image Deblurring Xintian Mao, Yiming Liu, Wei Shen, Qingli Li and Yan Wang code will be released soon

145 Dec 13, 2022
Lucid library adapted for PyTorch

Lucent PyTorch + Lucid = Lucent The wonderful Lucid library adapted for the wonderful PyTorch! Lucent is not affiliated with Lucid or OpenAI's Clarity

Lim Swee Kiat 520 Dec 26, 2022
Code for reproducing our paper: LMSOC: An Approach for Socially Sensitive Pretraining

LMSOC: An Approach for Socially Sensitive Pretraining Code for reproducing the paper LMSOC: An Approach for Socially Sensitive Pretraining to appear a

Twitter Research 11 Dec 20, 2022
The codebase for Data-driven general-purpose voice activity detection.

Data driven GPVAD Repository for the work in TASLP 2021 Voice activity detection in the wild: A data-driven approach using teacher-student training. S

Heinrich Dinkel 75 Nov 27, 2022
Alpha-IoU: A Family of Power Intersection over Union Losses for Bounding Box Regression

Alpha-IoU: A Family of Power Intersection over Union Losses for Bounding Box Regression YOLOv5 with alpha-IoU losses implemented in PyTorch. Example r

Jacobi(Jiabo He) 147 Dec 05, 2022
[AAAI 2022] Negative Sample Matters: A Renaissance of Metric Learning for Temporal Grounding

[AAAI 2022] Negative Sample Matters: A Renaissance of Metric Learning for Temporal Grounding Official Pytorch implementation of Negative Sample Matter

Multimedia Computing Group, Nanjing University 69 Dec 26, 2022
ERISHA is a mulitilingual multispeaker expressive speech synthesis framework. It can transfer the expressivity to the speaker's voice for which no expressive speech corpus is available.

ERISHA: Multilingual Multispeaker Expressive Text-to-Speech Library ERISHA is a multilingual multispeaker expressive speech synthesis framework. It ca

Ajinkya Kulkarni 43 Nov 27, 2022
a simple, efficient, and intuitive text editor

Oxygen beta a simple, efficient, and intuitive text editor Overview oxygen is a simple, efficient, and intuitive text editor designed as more featured

Aarush Gupta 1 Feb 23, 2022
Blind visual quality assessment on 360° Video based on progressive learning

Blind visual quality assessment on omnidirectional or 360 video (ProVQA) Blind VQA for 360° Video via Progressively Learning from Pixels, Frames and V

5 Jan 06, 2023
An implementation of MobileFormer

MobileFormer An implementation of MobileFormer proposed by Yinpeng Chen, Xiyang Dai et al. Including [1] Mobile-Former proposed in:

slwang9353 62 Dec 28, 2022
People movement type classifier with YOLOv4 detection and SORT tracking.

Movement classification The goal of this project would be movement classification of people, in other words, walking (normal and fast) and running. Yo

4 Sep 21, 2021