CSR---PDOS: Guiding and Validating Operator Behavior in Internet Services

互联网服务正迅速成为世界各地人们工作和休闲不可或缺的一部分。随之而来的是对24x7全天候可用性的需求。然而，最近的研究表明，互联网服务充其量只能达到99%-99.9%的可用性，这意味着每年有8到80个小时的停机时间。这个项目解决了不可用的一个主要来源：运营商的错误。具体地说，PI试图通过指导和验证运营商的行动来减少错误的影响，特别是侧重于基于集群的互联网服务。这种方法包括三个方面的努力：(1)通过采访和调查有经验的操作员，对志愿者操作员进行实验，以及举办几个操作员竞赛，来探索操作员错误的性质及其对互联网服务性能和可用性的影响；(2)开发可用于指导操作员基于出错可能性采取行动的操作员模型；以及(3)设计和原型验证基础设施，该验证基础设施是在线系统的一部分，但允许运营商在他们可以影响现场服务之前检查他们的行为的正确性。该项目的主要预期结果是示范，指导和验证运营商行动的系统可以显著减少运营商错误对服务可用性的影响。对于与可用性(和可管理性)相关的其他研究工作来说，有价值的人工制品将包括：(1)关于运营商错误及其对服务性能和可用性的影响的大量数据；(2)用于指导行动的运营商行为模型；以及(3)原型验证基础设施。

Internet services are rapidly becoming an integral part of people's work and leisure all over the world. With such popularity and importance comes the need for 24x7 availability. Yet, recent research suggests that, at best, Internet services are achieving only 99-99.9% availability, implying 8 to 80 hours of downtime per year.This project addresses a major source of unavailability: operator mistakes. Specifically, the PIs seek to reduce the impact of mistakes by guiding and validating operator actions, focusing specifically on cluster-based Internet services. This approach involves three efforts: (1) to explore the nature of operator mistakes and their impact on the performance and availability of Internet services by interviewing and surveying experienced operators, running experiments with volunteer operators, and running several operator contests; (2) to develop operator models that can be used to guide operator actions based on the likelihood of mistakes; and (3) to design and prototype a validation infrastructure that is part of the online system, yet allows operators to check the correctness of their actions before they can impact the live service.The main expected outcome of the project is a demonstration that systems that guide and validate operator actions can significantly reduce the impact of operator mistakes on service availability. Valuable artifacts for other research efforts relating to availability (and manageability) will include: (1) extensive data on operator mistakes and their impact on service performance and availability; (2) models of operator behavior for guiding actions; and (3) a prototype validation infrastructure.