论文部分内容阅读
当前的云计算系统,不论是虚拟化云还是分区云,难以同时满足用户体验和系统效率需求,产业界和学术界都开始研究下一代云计算系统以应对这个难题.本文指出,这个难题的一个重要原因是计算系统熵(无序、干扰和不确定性)居高不下,并归纳了云计算系统中存在的4类无序现象.本文提出了低熵云计算系统的学术概念,刻画了其主要特点,比较了低熵云计算系统与虚拟化云和分区云在用户体验、开发效率、运行效率、资源适配方面的区别,并讨论了低熵云的新概念和新技术:(1)不同于图灵可计算性和算法可计算性的实用可计算性概念,形式化地刻画了云计算行业的“用户体验差的功能是不存在的功能”的实践经验;(2)刻画云计算系统能够实现实用可计算性的充分必要条件,即DIP猜想;(3)支持DIP猜想,即能够区分、隔离、优先化计算任务相空间,从而降低干扰,有潜力同时满足用户体验和系统效率需求的标签化von Neumann体系结构;(4)适配深度学习负载与神经网络处理器的云计算协同设计技术.
The current cloud computing system, whether it is virtualized cloud or partitioned cloud, is difficult to satisfy both user experience and system efficiency requirements, and both industry and academia are beginning to study next-generation cloud computing system to cope with this problem.This paper points out that one of the problems The main reason is that the computing system entropy (disorder, interference and uncertainty) is high, and the four kinds of disordered phenomena in cloud computing system are summarized.This paper proposes the academic concept of low entropy cloud computing system, The main differences between low entropy cloud computing system and virtualized cloud and partitioned cloud in user experience, development efficiency, operational efficiency and resource adaptation are compared. New concepts and new technologies of low entropy cloud are discussed: (1) Different from the practical computability concepts of Turing computability and algorithmic computability, it formally depicts the practical experience of the cloud computing industry that “the function with poor user experience is a non-existent function”; (2) (3) support DIP conjecture, which can distinguish, isolate and prioritize computing task phase space, so as to reduce the interference and have potential at the same time Tagged von Neumann architecture to meet the needs of user experience and system efficiency; (4) cloud computing co-design technology to adapt deep learning load and neural network processor.