MiCS

Author:

Zhang Zhen1,Zheng Shuai2,Wang Yida2,Chiu Justin3,Karypis George2,Chilimbi Trishul3,Li Mu2,Jin Xin4

Affiliation:

1. Johns Hopkins University

2. Amazon Web Services

3. Amazon

4. Peking University

Abstract

Existing general purpose frameworks for gigantic model training, i.e., dense models with billions of parameters, cannot scale efficiently on cloud environment with various networking conditions due to large communication overheads. In this paper, we propose MiCS, which Minimizes the Communication Scale to bring down communication overhead. Specifically, by decreasing the number of participants in a communication collective, MiCS can utilize heterogeneous network bandwidth, reduce network traffic over slower links, reduce the latency of communications for maintaining high network bandwidth utilization, and amortize expensive global gradient synchronization overhead. Our evaluation on AWS shows that the system throughput of MiCS is up to 2.89× that of the state-of-the-art large model training systems. MiCS achieves near-linear scaling efficiency, which is up to 1.27× that of DeepSpeed. MiCS allows us to train a proprietary model with 100 billion parameters on 512 GPUs with 99.4% weak-scaling efficiency, and it is able to saturate over 54.5% theoretical computation power of each GPU on a public cloud with less GPU memory and more restricted networks than DGX-A100 clusters.

Publisher

Association for Computing Machinery (ACM)

Subject

General Earth and Planetary Sciences,Water Science and Technology,Geography, Planning and Development

Reference74 articles.

1. Optimization of MPI collective communication on BlueGene/L systems

2. Sanjith Athlur , Nitika Saran , Muthian Sivathanu , Ramachandran Ramjee , and Nipun Kwatra . 2021 . Varuna: Scalable, Low-cost Training of Massive Deep Learning Models. arXiv preprint arXiv:2111.04007 (2021). Sanjith Athlur, Nitika Saran, Muthian Sivathanu, Ramachandran Ramjee, and Nipun Kwatra. 2021. Varuna: Scalable, Low-cost Training of Massive Deep Learning Models. arXiv preprint arXiv:2111.04007 (2021).

3. AWS-P3-Instances 2022. Amazon EC2 P3 Instances. https://aws.amazon.com/ec2/instance-types/p3/. AWS-P3-Instances 2022. Amazon EC2 P3 Instances. https://aws.amazon.com/ec2/instance-types/p3/.

4. azure-gpu-ncv3-series 2022. Azure NCv3-series. https://docs.microsoft.com/en-us/azure/virtual-machines/ncv3-series. azure-gpu-ncv3-series 2022. Azure NCv3-series. https://docs.microsoft.com/en-us/azure/virtual-machines/ncv3-series.

5. azure-gpu-ndv2-series 2022. Azure Updated NDv2-series. https://docs.microsoft.com/en-us/azure/virtual-machines/ndv2-series. azure-gpu-ndv2-series 2022. Azure Updated NDv2-series. https://docs.microsoft.com/en-us/azure/virtual-machines/ndv2-series.

Cited by 5 articles. 订阅此论文施引文献 订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献

1. DynaPipe: Optimizing Multi-task Training through Dynamic Pipelines;Proceedings of the Nineteenth European Conference on Computer Systems;2024-04-22

2. GEMINI: Fast Failure Recovery in Distributed Training with In-Memory Checkpoints;Proceedings of the 29th Symposium on Operating Systems Principles;2023-10-23

3. Flash-LLM: Enabling Cost-Effective and Highly-Efficient Large Generative Model Inference with Unstructured Sparsity;Proceedings of the VLDB Endowment;2023-10

4. Training Large-scale Foundation Models on Emerging AI Chips;Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining;2023-08-04

5. xCCL: A Survey of Industry-Led Collective Communication Libraries for Deep Learning;Journal of Computer Science and Technology;2023-02

同舟云学术

1.学者识别学者识别

2.学术分析学术分析

3.人才评估人才评估

"同舟云学术"是以全球学者为主线,采集、加工和组织学术论文而形成的新型学术文献查询和分析系统,可以对全球学者进行文献检索和人才价值评估。用户可以通过关注某些学科领域的顶尖人物而持续追踪该领域的学科进展和研究前沿。经过近期的数据扩容,当前同舟云学术共收录了国内外主流学术期刊6万余种,收集的期刊论文及会议论文总量共计约1.5亿篇,并以每天添加12000余篇中外论文的速度递增。我们也可以为用户提供个性化、定制化的学者数据。欢迎来电咨询!咨询电话:010-8811{复制后删除}0370

www.globalauthorid.com

TOP

Copyright © 2019-2024 北京同舟云网络信息技术有限公司
京公网安备11010802033243号  京ICP备18003416号-3