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 is a JAX implementation of Neural Radiance Fields for learning purposes.

learn-nerf This is a JAX implementation of Neural Radiance Fields for learning purposes. I've been curious about NeRF and its follow-up work for a whi

Alex Nichol 62 Dec 20, 2022
Official implementation of Protected Attribute Suppression System, ICCV 2021

Official implementation of Protected Attribute Suppression System, ICCV 2021

Prithviraj Dhar 6 Jan 01, 2023
a grammar based feedback fuzzer

Nautilus NOTE: THIS IS AN OUTDATE REPOSITORY, THE CURRENT RELEASE IS AVAILABLE HERE. THIS REPO ONLY SERVES AS A REFERENCE FOR THE PAPER Nautilus is a

Chair for Sys­tems Se­cu­ri­ty 158 Dec 28, 2022
Code for Neurips2021 Paper "Topology-Imbalance Learning for Semi-Supervised Node Classification".

Topology-Imbalance Learning for Semi-Supervised Node Classification Introduction Code for NeurIPS 2021 paper "Topology-Imbalance Learning for Semi-Sup

Victor Chen 40 Nov 23, 2022
Neon-erc20-example - Example of creating SPL token and wrapping it with ERC20 interface in Neon EVM

Example of wrapping SPL token by ERC2-20 interface in Neon Requirements Install

7 Mar 28, 2022
NCVX (NonConVeX): A User-Friendly and Scalable Package for Nonconvex Optimization in Machine Learning.

NCVX NCVX: A User-Friendly and Scalable Package for Nonconvex Optimization in Machine Learning. Please check https://ncvx.org for detailed instruction

SUN Group @ UMN 28 Aug 03, 2022
Semi-supervised Semantic Segmentation with Directional Context-aware Consistency (CVPR 2021)

Semi-supervised Semantic Segmentation with Directional Context-aware Consistency (CAC) Xin Lai*, Zhuotao Tian*, Li Jiang, Shu Liu, Hengshuang Zhao, Li

Jia Research Lab 137 Dec 14, 2022
Exploiting a Zoo of Checkpoints for Unseen Tasks

Exploiting a Zoo of Checkpoints for Unseen Tasks This repo includes code to reproduce all results in the above Neurips paper, authored by Jiaji Huang,

Baidu Research 8 Sep 06, 2022
This is just a funny project that we want to see AutoEncoder (AE) can actually work to enhance the features we want

Funny_muscle_enhancer :) 1.Discription: This is just a funny project that we want to see AutoEncoder (AE) can actually work on the some features. We w

Jing-Yao Chen (Jacob) 8 Oct 01, 2022
BisQue is a web-based platform designed to provide researchers with organizational and quantitative analysis tools for 5D image data. Users can extend BisQue by implementing containerized ML workflows.

Overview BisQue is a web-based platform specifically designed to provide researchers with organizational and quantitative analysis tools for up to 5D

Vision Research Lab @ UCSB 26 Nov 29, 2022
Autotype on websites that have copy-paste disabled like Moodle, HackerEarth contest etc.

Autotype A quick and small python script that helps you autotype on websites that have copy paste disabled like Moodle, HackerEarth contests etc as it

Tushar 32 Nov 03, 2022
VACA: Designing Variational Graph Autoencoders for Interventional and Counterfactual Queries

VACA Code repository for the paper "VACA: Designing Variational Graph Autoencoders for Interventional and Counterfactual Queries (arXiv)". The impleme

Pablo Sánchez-Martín 16 Oct 10, 2022
[NeurIPS 2021] A weak-shot object detection approach by transferring semantic similarity and mask prior.

[NeurIPS 2021] A weak-shot object detection approach by transferring semantic similarity and mask prior.

BCMI 49 Jul 27, 2022
Adversarially Learned Inference

Adversarially Learned Inference Code for the Adversarially Learned Inference paper. Compiling the paper locally From the repo's root directory, $ cd p

Mohamed Ishmael Belghazi 308 Sep 24, 2022
Implicit MLE: Backpropagating Through Discrete Exponential Family Distributions

torch-imle Concise and self-contained PyTorch library implementing the I-MLE gradient estimator proposed in our NeurIPS 2021 paper Implicit MLE: Backp

UCL Natural Language Processing 249 Jan 03, 2023
Custom IMDB Dataset is extracted between 2020-2021 and custom distilBERT model is trained for movie success probability prediction

IMDB Success Predictor Project involves Web Scraping custom IMDB data between 2020 and 2021 of 10000 movies and shows sorted by number of votes ,fine

Gautam Diwan 1 Jan 18, 2022
CarND-LaneLines-P1 - Lane Finding Project for Self-Driving Car ND

Finding Lane Lines on the Road Overview When we drive, we use our eyes to decide where to go. The lines on the road that show us where the lanes are a

Udacity 769 Dec 27, 2022
This is an official implementation for "Exploiting Temporal Contexts with Strided Transformer for 3D Human Pose Estimation".

Exploiting Temporal Contexts with Strided Transformer for 3D Human Pose Estimation This repo is the official implementation of Exploiting Temporal Con

Vegetabird 241 Jan 07, 2023
[ICML 2021, Long Talk] Delving into Deep Imbalanced Regression

Delving into Deep Imbalanced Regression This repository contains the implementation code for paper: Delving into Deep Imbalanced Regression Yuzhe Yang

Yuzhe Yang 568 Dec 30, 2022