A Study of the Impact of Network Partitioning Faults on Cloud Systems

点击翻译按钮获取中文摘要

Google distributed systems are designed to be highly available despite failures of devices, machines, and**networks. Network-partitioning faults impose a unique challenge: They are some of the most**complex-to-handle faults but unfortunately are increasingly occurring in modern data centers.**We aim to conduct an in-depth characterization study of network-partitioning-related failures in five widely**used production-quality systems that mimic Google's internal systems. Our goal is to better understand the**impact of network partitioning on modern distributed systems, to understand the specific sequence of events**that lead to user-visible system failures, and to characterize these system failures to identify opportunities for**improving system fault tolerance. We plan to use our findings to design the network partitioning testing**(NPTest) framework, a system-level testing framework that can create and heal network partitions. For a wider**impact, we plan to make NPTest open source.**This project characterizes the impact of complex faults of modern networks and has the potential to have a**wide influence on technology, including Google Canada, which builds Web-scale systems. Furthermore, for**data centers (including the more than 200 centers located in Canada), this research program will improve the**reliability of a wide range of systems and bring benefits to a wide set of cloud applications, including**applications in fields such as medical informatics, online social networks, AI and machine learning systems,**smart energy distribution systems, and text analytics. Finally, this program will provide an invaluable**opportunity for training highly qualified personnel on cutting-edge networking and distributed systems**technology.