On the Design and Characteristics of Cloud Computing Systems Along the Performance-Failure Continuum

Cloud computing enables essential societal services, including communication, commerce, entertainment, and healthcare. It aims to be easy to use and low-cost so that as many fields as possible can reap its benefits. However, it is challenging to provide a highly performant and reliable service that is both easy to use and low-cost. We aim to help cloud system designers address this problem by providing a set of tools that enable reasoning about performance and reliability together. We posit that performance and reliability are part of the same performance-failure continuum. The techniques used to address low performance and unreliability share similarities. For example, when an RPC request is slow to respond, the application retries. The application also retries when a computation task fails in a data-processing workflow. Despite these similarities, the two concepts are studied separately. This separation leads to wasted work and even conflicting design objectives. We believe that high-performance and reliability techniques, when studied together, yield systems that better achieve their objectives. We provide conceptual and technical tools, as well as data, to enable this joint study. We provide conceptual and simulation tools to characterize cloud scheduler architectures and mechanisms in ExDe. We introduce the concept of scheduler frames to capture the underlying infrastructure that enables scheduling mechanisms and constrains the space of possible scheduler designs. We extend this work to scheduler APIs to characterize the impact of API design on scheduler performance. We characterize the performance-failure characteristics of two types of systems: hash-based task schedulers and serverless data store event trigger mechanisms. The studies capture the characteristics of systems under many scenarios by varying the workload, system load, and elastic behavior. We collect and analyze failure data for many cloud services across three categories: web, infrastructure cloud, and online games. We collect data from various vantage points, including user reports, operator reports, and online user counts. We demonstrate the use of this data to characterize system performance under realistic conditions, including failures, using trace-driven simulation.