从网格到“东数西算” : 构建国家算力基础设施

钱德沛; 栾钟治; 刘轶

doi:10.13700/j.bh.1001-5965.2022.0715

从网格到“东数西算” : 构建国家算力基础设施

doi: 10.13700/j.bh.1001-5965.2022.0715

北京航空航天大学计算机学院, 北京 100083

详细信息

通讯作者:
钱德沛, E-mail: depeiq@buaa.edu.cn

中图分类号: TP311;TP316;TP391
计量
- 文章访问数: 320
- HTML全文浏览量: 106
- PDF下载量: 61
- 被引次数: 0
出版历程
- 收稿日期: 2022-08-02
- 录用日期: 2022-08-18
- 网络出版日期: 2022-08-23

From grid to "East-west Computing Transfer" : Constructing national computing infrastructure

School of Computer Science and Engineering, Beihang University, Beijing 100083, China

More Information

Corresponding author: QIAN Depei, E-mail: depeiq@buaa.edu.cn

摘要

摘要:
简要回顾了几十年来计算机使用方式的变迁，介绍了基于网络计算技术的国家高性能计算基础设施CNGrid的设计与实现。讨论了在“东数西算”战略工程背景下中国算力发展的新趋势，以及国家算力基础设施发展面临的新的技术挑战，并对中国未来超算应用生态和算力基础设施建设提出了展望。
- 高性能计算 /
- 基础设施 /
- 网格计算 /
- CNGrid /
- 东数西算
Abstract:
This article gives a review of the evolution of computer use-mode over the time since the invention of modern computers and presents the challenges and tasks in building the national computing infrastructure. The first section of this article provides a brief overview of how computer use-mode has evolved over the past several decades. Then the design and implementation of China's national high performance computing infrastructure CNGrid are introduced. Following that, the trends and new technical challenges in developing the computing infrastructure under the circumstance of the national strategic project of "East-west Computing Transfer" are discussed. Finally, the perspectives of building the supercomputing eco-system and constructing the new type of computing infrastructure in China are presented.
- high performance computing /
- infrastructure /
- grid computing /
- CNGrid /
- East-west Computing Transfer

HTML全文

图 1 国家高性能计算基础设施的非集中层次虚拟化体系结构

Figure 1. Decentralized hierarchical virtualization architecture for national high performance computing infrastructure

下载: 全尺寸图片幻灯片

图 2 CNGrid Suite系统架构

Figure 2. CNGrid suite system architecture

下载: 全尺寸图片幻灯片

图 3 一体化基础设施监控管理体系结构框架

Figure 3. Architectural framework of integrated infrastructure monitoring and management

下载: 全尺寸图片幻灯片

图 4 工业社区系统框架

Figure 4. Industrial community system framework

下载: 全尺寸图片幻灯片

图 5 高性能计算应用集成开发环境系统架构

Figure 5. System architecture of integrated development environment for high performance computing applications

下载: 全尺寸图片幻灯片

表 1 世界部分网格相关研究计划

Table 1. Part of grid-related programs in the world

国别	网格计算相关研究计划/项目
美国	TeraGrid, XSEDE
欧盟	EGEE, EGI
英国	UK e-Science
日本	NAREGI，HPCI
韩国	K^*Grid
中国	CNGrid，ChinaGrid

下载: 导出CSV

表 2 中国科技部的网格和高性能计算项目

Table 2. Grid and high performance computing projects under the Ministry of Science and Technology of China

项目来源	项目名称	执行周期	项目成果
国家863重大课题	国家高性能计算环境	1999—2000年	4 000亿次曙光3000;包含5个高性能计算中心的国家高性能计算环境原型
国家863重大专项	高性能计算机及核心软件	2002—2005年	11.2万亿次曙光4000，5.36万亿次的联想深腾6800;国家高性能计算环境实验床“中国国家网格CNGrid”，8个结点，18万亿次计算能力; 一批网格应用
国家863重大项目	高效能计算机及网格服务环境	2006—2010年	4 700万亿次的天河1A，3 000万亿次的曙光6000，1 071万亿次的神威蓝光; 具有服务特征的国家网格服务环境CNGrid，11个结点，8 000万亿次计算能力; 一批网格和高性能计算应用
国家863重大项目	高效能计算机及应用服务环境	2011—2015年	12.5亿亿次的神威·太湖之光，10亿亿次的天河2A;以服务支持应用的国家高性能计算环境CNGrid，14个结点，20亿亿次计算能力; 一批高性能计算应用
国家重点研发专项	高性能计算	2016—2021年	E级计算机; 初步具备基础设施形态的国家高性能计算环境CNGrid，19个结点，52亿亿次计算能力; 一批高性能计算应用

下载: 导出CSV

表 3 典型深度神经网络模型的训练计算量^[25]

Table 3. Training computations of typical deep neural networks^[25]

神经网络模型	应用领域	训练计算量/Flops
AlexNet	图像分类	4.7×10¹⁷
VGG16	图像分类	8.5×10¹⁸
YOLOv3	图像目标检测	5.1×10¹⁹
Transformer	自然语言处理	7.4×10¹⁸
GPT-3	自然语言处理	3.1×10²³
注：数据来源于https://docs.google.com/spreadsheets/d/1AAIebjNsnJj_uKALHbXNfn3_YsT6sHXtCU0q7OIPuc4。

下载: 导出CSV

表 4 TOP500排名前十的高性能计算机(2022年6月)^[26]

Table 4. TOP10 in TOP500 high performance computing systems (June 2022)^[26]

排名	系统	处理器/加速器	Linpack性能/PFlops
1	Frontier	AMD 64C+AMD MI250X	1 102
2	Fugaku(富岳)	A64FX 48C	442.01
3	LUMI	AMD 64C+AMD MI250X	151.90
4	Summit(顶点)	IBM Power+Nvidia V100	148.60
5	Sierra(山脊)	IBM Power+Nvidia V100	94.64
6	Sunway TaihuLight(神威·太湖之光)	Sunway SW26010	93.01
7	Perlmutter	AMD 64C+Nvidia A100	70.87
8	Selene	AMD 64C+Nvidia A100	63.46
9	Tianhe-2A(天河2A)	Intel Xeon + Matrix2000	61.44
10	Adastra	AMD 64C+AMD MI250X	46.10
注：数据来源于http://www.top500.org。

下载: 导出CSV

参考文献(27)

[1]	DENNIS J. Segmentation and the design of multiprogrammed computer systems[J]Journal of the ACM, 1965, 12(4): 589-602. doi: 10.1145/321296.321310
[2]	SACKMAN H. Time-sharing versus batch processing: The experimental evidence[C]//Proceedings of the American Federation of Information Processing Societies. New York: ACM, 1968: 1-10.
[3]	SCHWARTZ J, COFFMAN E, WEISSMAN C. A general-purpose time-sharing system[C]//Proceedings of the American Federation of Information Processing Societies. New York: ACM, 1964: 397-411.
[4]	MILLS D L, BRAUN H. The NSFNET backbone network[C]//Proceedings of the ACM Workshop on Frontiers in Computer Communications Technology. New York: ACM, 1987: 191-196.
[5]	FOSTER I T, KESSELMAN C. The grid: Blueprint for a new computing infrastructure[M]. San Francisco: Morgan Kaufman Publishers, 1998.
[6]	STEVENS R, WOODWARD P, DEFANTI T, et al. From the I-WAY to the national technology grid[J]. Communications of the ACM, 1997, 40(11): 50-60. doi: 10.1145/265684.265692
[7]	THOMAS M, BOISSEAU J, DAHAN M, et al. Development of NPACI grid application portals and portal Web services[J]. Cluster Computing, 2003, 6(3): 177-188. doi: 10.1023/A:1023566402391
[8]	FOSTER I, CZAJKOWSKI K, FERGUSON D, et al. Modeling and managing state in distributed systems: The role of OGSI and WSRF[J]. Proceedings of the IEEE, 2005, 93(3): 604-612. doi: 10.1109/JPROC.2004.842766
[9]	TALIA D. The open grid services architecture: Where the grid meets the Web[J]. IEEE Internet Computing, 2002, 6(6): 67-71. doi: 10.1109/MIC.2002.1067739
[10]	FOSTER I, KESSELMAN C. Globus: A metacomputing infrastructure toolkit[J]. International Journal of Supercomputer Application, 1998, 11(2): 115-129.
[11]	REED D.A. Grids, the TeraGrid, and beyond[J]. IEEE Computer, 2003, 36(1): 62-68. doi: 10.1109/MC.2003.1160057
[12]	KUNSZT P. European DataGrid project: Status and plans[J]. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 2003, 502(2-3): 376-381. doi: 10.1016/S0168-9002(03)00447-9
[13]	GAGLIARDI F, JONES B, GREY F, et al. Building an infrastructure for scientific grid computing: Status and goals of the EGEE project[J]. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 2005, 363(1833): 1729-1742. doi: 10.1098/rsta.2005.1603
[14]	HEY T, TREFETHEN A E. The UK e-Science core programme and the grid[J]. Future Generation Computer Systems, 2002, 18(8): 1017-1031. doi: 10.1016/S0167-739X(02)00082-1
[15]	MATSUOKA S, SHINJO S, AOYAGI M, et al. Japanese computational grid research project: NAREGI[J]. Proceedings of the IEEE, 2005, 93(3): 522-533. doi: 10.1109/JPROC.2004.842748
[16]	ARMBRUST M, FOX A, GRIFFITH R, et al. Above the clouds: A Berkeley view of cloud computing: UCB/EECS-2009-28[R]. Berkeley: EECS Department University of California, Berkeley Technical Report, 2009.
[17]	SARASWAT M, TRIPATHI R C. Cloud computing: Analysis of top 5 CSPs in SaaS, PaaS and IaaS platforms[C]//2020 9th International Conference on System Modeling and Advancement in Research Trends, 2020: 20421390.
[18]	SOTOMAYOR B, MONTERO R, LLORENTE I, et al. Virtual infrastructure management in private and hybrid clouds[J]. IEEE Internet Computing, 2009, 13(5): 14-22. doi: 10.1109/MIC.2009.119
[19]	BARIK R, LENKA R, RAO K, et al. Performance analysis of virtual machines and containers in cloud computing[C]//2016 International Conference on Computing, Communication and Automation. Piscataway: IEEE Press, 2016: 16585534.
[20]	SIMONS J. HPC cloud bad; HPC in the cloud good[C]//2013 IEEE 27th International Symposium on Parallel and Distributed Processing. Piscataway: IEEE Press, 2013: 13683523.
[21]	MOR N. Edge computing: Scaling resources within multiple administrative domains[J]. Queue, 2018, 16(6): 106-116. doi: 10.1145/3305263.3313377
[22]	乔健, 查礼. 中国国家网格作业管理设计与实现[J]. 计算机应用, 2008, 28(8): 2003-2009. https://www.cnki.com.cn/Article/CJFDTOTAL-JSJY200808030.htm QIAO J, ZHA L. Design and implementation of grid job management for China national grid[J]. Computer Applications, 2008, 28(8): 2003-2009(in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-JSJY200808030.htm
[23]	王小宁, 肖海力, 曹荣强. 面向高性能计算环境的作业优化调度模型的设计与实现[J]. 计算机工程与科学, 2017, 39(4): 619-626. doi: 10.3969/j.issn.1007-130X.2017.04.002 WANG X N, XIAO H L, CAO R Q. Design and implementation of an optimal job scheduling model for the high performance computing environment[J]. Computer Engineering & Science, 2017, 39(4): 619-626(in Chinese). doi: 10.3969/j.issn.1007-130X.2017.04.002
[24]	喻林, 邹永强, 查礼. CNGrid GOS安全: 设计与实现[J]. 华中科技大学学报(自然科学版), 2010, 38(S1): 6-10. https://www.cnki.com.cn/Article/CJFDTOTAL-HZLG2010S1003.htm YU L, ZOU Y Q, ZHA L. CNGrid GOS security: Design and implementation[J]. Journal of Huazhong University of Science & Technology (Natural Science Edition), 2010, 38(S1): 6-10 (in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-HZLG2010S1003.htm
[25]	SEVILLA J, VILLALOBOS P, C ERON J, et al. Parameter, compute and data trends in machine learning[EB/OL]. [2022-05-30]. https://docs.google.com/spreadsheets/d/1AAIebj NsnJj_uKALHbXNfn3_YsT6sHXtCU0q7OIPuc4.
[26]	TOP500 list[EB/OL]. [2022-06-20]. https://top500.org/lists/top500/2022/06/.
[27]	BONAWITZ K, EICHNER H, GRIESKAMP W, et al. Towards federated learning at scale: System design[C]//Proceedings of the Conference on Machine Learning and Systems. Piscataway: IEEE Press, 2019: 1-15.