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。。。
VGGVox models for Speaker Identification and Verification trained on the VoxCeleb (1 & 2) datasets

VGGVox models for speaker identification and verification This directory contains code to import and evaluate the speaker identification and verificat

338 Dec 27, 2022
Image Recognition using Pytorch

PyTorch Project Template A simple and well designed structure is essential for any Deep Learning project, so after a lot practice and contributing in

Sarat Chinni 1 Nov 02, 2021
Hide screen when boss is approaching.

BossSensor Hide your screen when your boss is approaching. Demo The boss stands up. He is approaching. When he is approaching, the program fetches fac

Hiroki Nakayama 6.2k Jan 07, 2023
This is the latest version of the PULP SDK

PULP-SDK This is the latest version of the PULP SDK, which is under active development. The previous (now legacy) version, which is no longer supporte

78 Dec 07, 2022
[TPDS'21] COSCO: Container Orchestration using Co-Simulation and Gradient Based Optimization for Fog Computing Environments

COSCO Framework COSCO is an AI based coupled-simulation and container orchestration framework for integrated Edge, Fog and Cloud Computing Environment

imperial-qore 39 Dec 25, 2022
SigOpt wrappers for scikit-learn methods

SigOpt + scikit-learn Interfacing This package implements useful interfaces and wrappers for using SigOpt and scikit-learn together Getting Started In

SigOpt 73 Sep 30, 2022
Pytorch implementation of One-Shot Affordance Detection

One-shot Affordance Detection PyTorch implementation of our one-shot affordance detection models. This repository contains PyTorch evaluation code, tr

46 Dec 12, 2022
PyTorch CZSL framework containing GQA, the open-world setting, and the CGE and CompCos methods.

Compositional Zero-Shot Learning This is the official PyTorch code of the CVPR 2021 works Learning Graph Embeddings for Compositional Zero-shot Learni

EML Tübingen 70 Dec 27, 2022
Dynamics-aware Adversarial Attack of 3D Sparse Convolution Network

Leaded Gradient Method (LGM) This repository contains the PyTorch implementation for paper Dynamics-aware Adversarial Attack of 3D Sparse Convolution

An Tao 2 Oct 18, 2022
A knowledge base construction engine for richly formatted data

Fonduer is a Python package and framework for building knowledge base construction (KBC) applications from richly formatted data. Note that Fonduer is

HazyResearch 386 Dec 05, 2022
wmctrl ported to Python Ctypes

work in progress wmctrl is a command that can be used to interact with an X Window manager that is compatible with the EWMH/NetWM specification. wmctr

Iyad Ahmed 22 Dec 31, 2022
A benchmark dataset for mesh multi-label-classification based on cube engravings introduced in MeshCNN

Double Cube Engravings This script creates a dataset for multi-label mesh clasification, with an intentionally difficult setup for point cloud classif

Yotam Erel 1 Nov 30, 2021
CVPR 2021 Official Pytorch Code for UC2: Universal Cross-lingual Cross-modal Vision-and-Language Pre-training

UC2 UC2: Universal Cross-lingual Cross-modal Vision-and-Language Pre-training Mingyang Zhou, Luowei Zhou, Shuohang Wang, Yu Cheng, Linjie Li, Zhou Yu,

Mingyang Zhou 28 Dec 30, 2022
Training Certifiably Robust Neural Networks with Efficient Local Lipschitz Bounds (Local-Lip)

Training Certifiably Robust Neural Networks with Efficient Local Lipschitz Bounds (Local-Lip) Introduction TL;DR: We propose an efficient and trainabl

17 Dec 01, 2022
PyTorch implementation for the paper Visual Representation Learning with Self-Supervised Attention for Low-Label High-Data Regime

Visual Representation Learning with Self-Supervised Attention for Low-Label High-Data Regime Created by Prarthana Bhattacharyya. Disclaimer: This is n

Prarthana Bhattacharyya 5 Nov 08, 2022
Scripts used to make and evaluate OpenAlex's concept tagging model

openalex-concept-tagging This repository contains all of the code for getting the concept tagger up and running. To learn more about where this model

OurResearch 18 Dec 09, 2022
Immortal tracker

Immortal_tracker Prerequisite Our code is tested for Python 3.6. To install required liabraries: pip install -r requirements.txt Waymo Open Dataset P

74 Dec 03, 2022
PyTorch implementation of "Transparency by Design: Closing the Gap Between Performance and Interpretability in Visual Reasoning"

Transparency-by-Design networks (TbD-nets) This repository contains code for replicating the experiments and visualizations from the paper Transparenc

David Mascharka 351 Nov 18, 2022
A package related to building quasi-fibration symmetries

qf A package related to building quasi-fibration symmetries. If you'd like to learn more about how it works, see the brief explanation and References

Paolo Boldi 1 Dec 01, 2021
Torch implementation of various types of GAN (e.g. DCGAN, ALI, Context-encoder, DiscoGAN, CycleGAN, EBGAN, LSGAN)

gans-collection.torch Torch implementation of various types of GANs (e.g. DCGAN, ALI, Context-encoder, DiscoGAN, CycleGAN, EBGAN). Note that EBGAN and

Minchul Shin 53 Jan 22, 2022