高性能计算多层次不连续非线性可扩展现象研究  

Multi-Level Discontinuous and Nonlinear Scalability Phenomenon in High Performance Computing

作　　者：张云泉[1] 袁良[1] 陈一峯 冯晓兵[1] 张贺[3] ZHANG Yun-Quan;YUAN Liang;CHEN Yi-Feng;FENG Xiao-Bing;ZHANG He(State Key Laboratory of Computer Architecture,Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190;Peking University,Beijing 100871;Institute of Atmospheric Physics,Chinese Academy of Sciences,Beijing 100029)

机构地区：[1]中国科学院计算技术研究所计算机体系结构国家重点实验室,北京100190 [2]北京大学,北京100871 [3]中国科学院大气物理研究所,北京100029

出　　处：《计算机学报》2020年第6期973-989,共17页Chinese Journal of Computers

基　　金：国家重点研发计划(2016YFB0200803);国家自然科学基金(61432018,61521092,61272136,61402441,61502450)资助.

摘　　要：高性能计算是计算科学的具体实践,极大地促进了各领域的科学进展,也对国家的经济建设起到了无法替代的基础性作用.从几十年发展的时间尺度和十万至百万核量级并行规模尺度研究大规模并行软件的研制发展历史来看,发现大规模并行应用软件的开发中物理模型、并行算法、并行软件实现以及底层硬件多个层次中存在的可扩展性的两种有趣现象,即不连续性和非线性现象.本文总结分析这一普遍存在现象,系统梳理计算机软硬件发展,特别是高性能计算发展中的可扩展问题,为未来并行计算领域发展提供方法论层面的借鉴和指导.High Performance Computing(HPC)is the practice of computational science.It promotes the scienctific progress of many fields and serves as an irreplaceable foundation for national economic development.From the perspective of the large-scale parallel software development with a timescale of decades and the scalabilities of thousands and millions cores,we observe two interesting phenomena in multi-level of HPC including physical models,parallel algorithms,parallel software implementations and underlying hardware architectures:discountinuous and nonlinear.This paper presents a detailed study of this phenomenon especially the scalability development in HPC.This paper also provides a methodology guidance for future parallel computing progress.

关 键 词：高性能计算 超级计算 可扩展性 多层次 不连续 非线性 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]