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。。。
PromptDet: Expand Your Detector Vocabulary with Uncurated Images

PromptDet: Expand Your Detector Vocabulary with Uncurated Images Paper Website Introduction The goal of this work is to establish a scalable pipeline

103 Dec 20, 2022
Use .csv files to record, play and evaluate motion capture data.

Purpose These scripts allow you to record mocap data to, and play from .csv files. This approach facilitates parsing of body movement data in statisti

21 Dec 12, 2022
Implementation of the GVP-Transformer, which was used in the paper "Learning inverse folding from millions of predicted structures" for de novo protein design alongside Alphafold2

GVP Transformer (wip) Implementation of the GVP-Transformer, which was used in the paper Learning inverse folding from millions of predicted structure

Phil Wang 19 May 06, 2022
Code repository for EMNLP 2021 paper 'Adversarial Attacks on Knowledge Graph Embeddings via Instance Attribution Methods'

Adversarial Attacks on Knowledge Graph Embeddings via Instance Attribution Methods This is the code repository to accompany the EMNLP 2021 paper on ad

Peru Bhardwaj 7 Sep 25, 2022
Efficiently Disentangle Causal Representations

Efficiently Disentangle Causal Representations Install dependency pip install -r requirements.txt Main experiments Causality direction prediction cd

4 Apr 01, 2022
Code for Universal Semi-Supervised Semantic Segmentation models paper accepted in ICCV 2019

USSS_ICCV19 Code for Universal Semi Supervised Semantic Segmentation accepted to ICCV 2019. Full Paper available at https://arxiv.org/abs/1811.10323.

Tarun K 68 Nov 24, 2022
Implementation of NÜWA, state of the art attention network for text to video synthesis, in Pytorch

NÜWA - Pytorch (wip) Implementation of NÜWA, state of the art attention network for text to video synthesis, in Pytorch. This repository will be popul

Phil Wang 463 Dec 28, 2022
This reposityory contains the PyTorch implementation of our paper "Generative Dynamic Patch Attack".

Generative Dynamic Patch Attack This reposityory contains the PyTorch implementation of our paper "Generative Dynamic Patch Attack". Requirements PyTo

Xiang Li 8 Nov 17, 2022
Latte: Cross-framework Python Package for Evaluation of Latent-based Generative Models

Cross-framework Python Package for Evaluation of Latent-based Generative Models Latte Latte (for LATent Tensor Evaluation) is a cross-framework Python

Karn Watcharasupat 30 Sep 08, 2022
Code for "Typilus: Neural Type Hints" PLDI 2020

Typilus A deep learning algorithm for predicting types in Python. Please find a preprint here. This repository contains its implementation (src/) and

47 Nov 08, 2022
A DCGAN to generate anime faces using custom mined dataset

Anime-Face-GAN-Keras A DCGAN to generate anime faces using custom dataset in Keras. Dataset The dataset is created by crawling anime database websites

Pavitrakumar P 190 Jan 03, 2023
Scalable Multi-Agent Reinforcement Learning

Scalable Multi-Agent Reinforcement Learning 1. Featured algorithms: Value Function Factorization with Variable Agent Sub-Teams (VAST) [1] 2. Implement

3 Aug 02, 2022
E-RAFT: Dense Optical Flow from Event Cameras

E-RAFT: Dense Optical Flow from Event Cameras This is the code for the paper E-RAFT: Dense Optical Flow from Event Cameras by Mathias Gehrig, Mario Mi

Robotics and Perception Group 71 Dec 12, 2022
A LiDAR point cloud cluster for panoptic segmentation

Divide-and-Merge-LiDAR-Panoptic-Cluster A demo video of our method with semantic prior: More information will be coming soon! As a PhD student, I don'

YimingZhao 65 Dec 22, 2022
Making Structure-from-Motion (COLMAP) more robust to symmetries and duplicated structures

SfM disambiguation with COLMAP About Structure-from-Motion generally fails when the scene exhibits symmetries and duplicated structures. In this repos

Computer Vision and Geometry Lab 193 Dec 26, 2022
Deep Reinforced Attention Regression for Partial Sketch Based Image Retrieval.

DARP-SBIR Intro This repository contains the source code implementation for ICDM submission paper Deep Reinforced Attention Regression for Partial Ske

2 Jan 09, 2022
Chinese license plate recognition

AgentCLPR 简介 一个基于 ONNXRuntime、AgentOCR 和 License-Plate-Detector 项目开发的中国车牌检测识别系统。 车牌识别效果 支持多种车牌的检测和识别(其中单层车牌识别效果较好): 单层车牌: [[[[373, 282], [69, 284],

AgentMaker 26 Dec 25, 2022
A motion detection system with RaspberryPi, OpenCV, Python

Human Detection System using Raspberry Pi Functionality Activates a relay on detecting motion. You may need following components to get the expected R

Omal Perera 55 Dec 04, 2022
Mail classification with tensorflow and MS Exchange Server (ham or spam).

Mail classification with tensorflow and MS Exchange Server (ham or spam).

Metin Karatas 1 Sep 11, 2021
SAGE: Sensitivity-guided Adaptive Learning Rate for Transformers

SAGE: Sensitivity-guided Adaptive Learning Rate for Transformers This repo contains our codes for the paper "No Parameters Left Behind: Sensitivity Gu

Chen Liang 23 Nov 07, 2022