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。。。
Semantic Segmentation of images using PixelLib with help of Pascalvoc dataset trained with Deeplabv3+ framework.

CARscan- Approach 1 - Segmentation of images by detecting contours. It failed because in images with elements along with cars were also getting detect

Padmanabha Banerjee 5 Jul 29, 2021
Official PyTorch implementation of the paper "Self-Supervised Relational Reasoning for Representation Learning", NeurIPS 2020 Spotlight.

Official PyTorch implementation of the paper: "Self-Supervised Relational Reasoning for Representation Learning" (2020), Patacchiola, M., and Storkey,

Massimiliano Patacchiola 135 Jan 03, 2023
SegTransVAE: Hybrid CNN - Transformer with Regularization for medical image segmentation

SegTransVAE: Hybrid CNN - Transformer with Regularization for medical image segmentation This repo is the official implementation for SegTransVAE. Seg

Nguyen Truong Hai 4 Aug 04, 2022
An unopinionated replacement for PyTorch's Dataset and ImageFolder, that handles Tar archives

Simple Tar Dataset An unopinionated replacement for PyTorch's Dataset and ImageFolder classes, for datasets stored as uncompressed Tar archives. Just

Joao Henriques 47 Dec 20, 2022
PyTorch and GPyTorch implementation of the paper "Conditioning Sparse Variational Gaussian Processes for Online Decision-making."

Conditioning Sparse Variational Gaussian Processes for Online Decision-making This repository contains a PyTorch and GPyTorch implementation of the pa

Wesley Maddox 16 Dec 08, 2022
This is the repository for our paper SimpleTrack: Understanding and Rethinking 3D Multi-object Tracking

SimpleTrack This is the repository for our paper SimpleTrack: Understanding and Rethinking 3D Multi-object Tracking. We are still working on writing t

TuSimple 189 Dec 26, 2022
Experiments for Fake News explainability project

fake-news-explainability Experiments for fake news explainability project This repository only contains the notebooks used to train the models and eva

Lorenzo Flores (Lj) 1 Dec 03, 2022
Generalizing Gaze Estimation with Outlier-guided Collaborative Adaptation

Generalizing Gaze Estimation with Outlier-guided Collaborative Adaptation Our paper is accepted by ICCV2021. Picture: Overview of the proposed Plug-an

Yunfei Liu 32 Dec 10, 2022
VR Viewport Pose Model for Quantifying and Exploiting Frame Correlations

This repository contains the introduction to the collected VRViewportPose dataset and the code for the IEEE INFOCOM 2022 paper: "VR Viewport Pose Model for Quantifying and Exploiting Frame Correlatio

0 Aug 10, 2022
Easy-to-use micro-wrappers for Gym and PettingZoo based RL Environments

SuperSuit introduces a collection of small functions which can wrap reinforcement learning environments to do preprocessing ('microwrappers'). We supp

Farama Foundation 357 Jan 06, 2023
Multi-objective constrained optimization for energy applications via tree ensembles

Multi-objective constrained optimization for energy applications via tree ensembles

C⚙G - Imperial College London 1 Nov 19, 2021
Weakly Supervised Segmentation by Tensorflow.

Weakly Supervised Segmentation by Tensorflow. Implements semantic segmentation in Simple Does It: Weakly Supervised Instance and Semantic Segmentation, by Khoreva et al. (CVPR 2017).

CHENG-YOU LU 52 Dec 27, 2022
CLIP + VQGAN / PixelDraw

clipit Yet Another VQGAN-CLIP Codebase This started as a fork of @nerdyrodent's VQGAN-CLIP code which was based on the notebooks of @RiversWithWings a

dribnet 276 Dec 12, 2022
百度2021年语言与智能技术竞赛机器阅读理解Pytorch版baseline

项目说明: 百度2021年语言与智能技术竞赛机器阅读理解Pytorch版baseline 比赛链接:https://aistudio.baidu.com/aistudio/competition/detail/66?isFromLuge=true 官方的baseline版本是基于paddlepadd

周俊贤 54 Nov 23, 2022
Code repo for "Transformer on a Diet" paper

Transformer on a Diet Reference: C Wang, Z Ye, A Zhang, Z Zhang, A Smola. "Transformer on a Diet". arXiv preprint arXiv (2020). Installation pip insta

cgraywang 31 Sep 26, 2021
🎃 Core identification module of AI powerful point reading system platform.

ppReader-Kernel Intro Core identification module of AI powerful point reading system platform. Usage 硬件: Windows10、GPU:nvdia GTX 1060 、普通RBG相机 软件: con

CrashKing 1 Jan 11, 2022
Pytorch implementation of Learning with Opponent-Learning Awareness

Pytorch implementation of Learning with Opponent-Learning Awareness using DiCE

Alexis David Jacq 82 Sep 15, 2022
A PaddlePaddle implementation of Time Interval Aware Self-Attentive Sequential Recommendation.

TiSASRec.paddle A PaddlePaddle implementation of Time Interval Aware Self-Attentive Sequential Recommendation. Introduction 论文:Time Interval Aware Sel

Paddorch 2 Nov 28, 2021
How to use TensorLayer

How to use TensorLayer While research in Deep Learning continues to improve the world, we use a bunch of tricks to implement algorithms with TensorLay

zhangrui 349 Dec 07, 2022
The implemention of Video Depth Estimation by Fusing Flow-to-Depth Proposals

Flow-to-depth (FDNet) video-depth-estimation This is the implementation of paper Video Depth Estimation by Fusing Flow-to-Depth Proposals Jiaxin Xie,

32 Jun 14, 2022