Low-entropy cloud computing systems

Abstract

Current cloud computing systems, whether virtualization clouds or partitioned clouds, face the challenge of simultaneously satisfying user experience and system efficiency requirements. Both the industry and the academia are investigating next-generation cloud computing systems to address this problem. This paper points out a main cause of this problem: existing cloud systems have high computing system entropy (i.e., disorder and uncertainty), which manifest as four classes of disorders. We propose a new concept of “low-entropy cloud computing systems, and contrast them to virtualization clouds and partitioned clouds, in terms of user experience, application development efficiency, execution efficiency, and resource matching. We discuss four new features and techniques of low-entropy clouds: (1) a notion of production computability that, unlike Turing computability and algorithmic tractability, formalizes the user experience requirements of cloud computing practices; (2) a conjecture, named the DIP (differentiation, isolation, prioritization) conjecture, that tries to capture the necessary and sufficient conditions for a cloud computing system to realize production computability; (3) the labeled von Neumann architecture that has the potential to support the DIP capabilities and thus simultaneously satisfy user experience and system efficiency requirements; and (4) a co-design technique allowing a cloud computing system to adaptively match deep-learning workloads to neural network accelerator hardware.