Optimizing Network Performance for Distributed DNN Training on GPU Clusters: ImageNet/AlexNet Training in 1.5 Minutes

By Peng Sun, Wansen Feng, Ruobing Han, Shengen Yan, Yonggang Wen. 2019

72 citations recorded by the source corpus.

It is important to scale out deep neural network (DNN) training for reducing model training time. The high communication overhead is one of the major performance bottlenecks for distributed DNN training across multiple GPUs. Our investigations have shown that popular open-source DNN systems could only achieve 2.5 speedup ratio on 64 GPUs connected by 56 Gbps network. To address this problem, we propose a communication backend named GradientFlow for distributed DNN training, and employ a set of network optimization techniques. First, we integrate ring-based allreduce, mixed-precision training, and computation/communication overlap into GradientFlow. Second, we propose lazy allreduce to improve network throughput by fusing multiple communication operations into a single one, and design coarse-grained sparse communication to reduce network traffic by only transmitting important gradient chunks. When training ImageNet/AlexNet on 512 GPUs, our approach achieves 410.2 speedup ratio and completes 95-epoch training in 1.5 minutes, which outperforms existing approaches.

Subfields: Computer Vision and Pattern Recognition, Neurology, Radiology, Nuclear Medicine and Imaging

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

1. Carole-Jean Wu
2. David M. Brooks
3. Gu-Yeon Wei
4. Udit Gupta
5. Matei A. Zaharia
6. D. Narayanan
7. Dehao Chen
8. David A. Patterson

In Neurology

1. Khaled Ben Letaief
2. Yuanming Shi
3. Debo Dutta

In Radiology, Nuclear Medicine and Imaging

1. Khaled Ben Letaief
2. Yuanming Shi
3. Hanlin Tang

Citations in the corpus

• MLPerf Training Benchmark (2019, background): “ResNet-50 is a residual network (He et al., 2016a;b); such networks and their derivatives remain the state of the art in image classification, and system studies commonly use them (Goyal et al., 2017; Jia et al., 2018; Mikami et al., 2018; Ying et al., 2018; Sun et al., 2019).”
• Communication-Efficient Edge AI: Algorithms and Systems (2020, methodology): “With 512 Volta GPUs interconnected at the rate of 56Gbps, the Alexnet can be trained within record of 2.6 minutes in the data center of SenseTime [31].”