Optimizing Distributed DNN Training Using CPUs and BlueField-2 DPUs

The deep learning (DL) training process consists of multiple phases—data augmentation, training, and validation of the trained model. Traditionally, these phases are executed either on the central processing units or graphics processing units in a serial fashion due to lack of additional computing r...

Full description

Saved in:
Bibliographic Details
Published in:IEEE MICRO 2022-03, Vol.42 (2), p.53-60
Main Authors: Jain, Arpan, Alnaasan, Nawras, Shafi, Aamir, Subramoni, Hari, Panda, Dhabaleswar K.
Format: Article
Language:English
Subjects:
ARM
Central processing units
Circuits
Computational modeling
CPU
CPUs
Data models
Data processing
Deep learning
DNN Training
DPU
Graphics processing units
Horovod
MPI
Network adapters
Neural networks
Offloading
Parallel processing
Phases
Program processors
PyTorch
Training
Training data
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The deep learning (DL) training process consists of multiple phases—data augmentation, training, and validation of the trained model. Traditionally, these phases are executed either on the central processing units or graphics processing units in a serial fashion due to lack of additional computing resources to offload independent phases of DL training. Recently, Mellanox/NVIDIA introduced the BlueField-2 data processing units (DPUs), which combine the advanced capabilities of traditional application-specific-integrated-circuit-based network adapters with an array of ARM processors. In this article, we explore how to take advantage of the additional ARM cores on the BlueField-2 DPUs. We propose and evaluate multiple novel designs to efficiently offload the phases of DL training to the DPUs. Our experimental results show that the proposed designs are able to deliver up to 17.5% improvement in overall DL training time. To the best of our knowledge, this is the first work to explore the use of DPUs to accelerate DL training.
ISSN:0272-1732
1937-4143
DOI:10.1109/MM.2021.3139027