ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2020, Vol. 57 ›› Issue (5): 897-905.doi: 10.7544/issn1000-1239.2020.20200058

• 综述 •    下一篇

算礼:探索计算系统的可分析抽象

徐志伟,王一帆,赵永威,李春典   

  1. (计算机体系结构国家重点实验室(中国科学院计算技术研究所) 北京 100190) (中国科学院大学 北京 100049) (zxu@ict.ac.cn)
  • 出版日期: 2020-05-01
  • 基金资助: 
    国家重点研发计划项目(2016YFB1000200);国家自然科学基金重点项目(61532016);中国科学院网络计算创新研究院物端计算系统项目

Computation Protocols: Analyzable Abstractions for Computing Systems

Xu Zhiwei, Wang Yifan, Zhao Yongwei, Li Chundian   

  1. (State Key Laboratory of Computer Architecture (Institute of Computing Technology, Chinese Academy of Sciences), Beijing 100190) (University of Chinese Academy of Sciences, Beijing 100049)
  • Online: 2020-05-01
  • Supported by: 
    This work was supported by the National Key Research and Development Program of China (2016YFB1000200), the Key Program of the National Natural Science Foundation of China (61532016), and the Things Computing System Project of CAS Network Computing Innovation Institute.

摘要: 计算机系统结构研究正在进入多样性时代.同时,以原型系统构建和基准程序测试为主要特征的计算系统研究方法,使得计算系统的研究成本依然居高不下,难以应对多样性挑战.这个矛盾呼唤新的可分析计算系统学术抽象,其主要特征是研究某个新系统时,在原型系统实现和基准程序测试之前,就能够分析出该系统的主要性质,进而筛掉不合适的候选系统,大幅度降低研究成本.这正是作为计算机应用抽象的算法概念所具有的特征:在算法实现和基准测试之前就基本可以分析出该算法的时间复杂度和空间复杂度等主要性质.首先,归纳了算法抽象的7条优点,指出最值得计算系统研究学习的是可分析抽象.其次,回顾了系统抽象的相关工作和历史经验,并提出了一个初步候选,称为算礼(computation protocol).最后,讨论了算礼的通用定义、黑箱表示和白箱表示,并用初步的实例指出,算礼思想有助于在计算系统领域提出系统猜想、分析新的并行计算模型、拓展现有架构、启发新的系统评价方法.

关键词: 算礼, 可分析抽象, 计算系统研究, 原型系统实现, 基准程序测试

Abstract: Computing systems research is entering an era of diversity. At the same time, systems research still mainly follows the prototype development and benchmark evaluation approach, making the research cost too high to address the diversity challenge. This dilemma calls for new analyzable abstractions of computing systems. When researching a new system, we can use its abstraction to analyze its characteristics to filter out inappropriate candidate systems before costly prototyping and benchmarking. We already have such a concept for computer applications, called algorithm. Before an algorithm’s implementation and benchmark evaluation, we can usually analyze its main properties, such as time complexity and space complexity. In this paper, we summarize seven advantages of the algorithm concept and propose a preliminary counterpart for computing systems, called computation protocol. Learning from six historical lessons from systems research, we discuss a general definition, a black-box representation, and a white-box representation of the computation protocol concept. We use preliminary examples to point out that computation protocol thinking may be helpful to propose computing systems conjecture, analyze new parallel computing model, extend existing systems architecture, and inspire new system evaluation method.

Key words: computation protocol, analyzable abstraction, computing systems research, prototype system development, benchmark evaluation

中图分类号: