Scalable Self-Healing Systems

今天的企业信息系统包括软件、计算机、网络等的大型、复杂的相互连接的集合。这些系统是任务关键型的，因此期望具有高可用性、提供良好的性能并在有限的预算内运行。目前，这是通过让高度可用的数据中心管理员主动监视关键服务、识别故障、违反服务级别协议(SLA)等、定位和诊断故障组件以及总体上确保系统处于正常运行状态来实现的。这种方法成本高昂(美国每年的商业成本约为1,000亿美元，到2018年增长30%；美国劳工统计局)。虽然最近的研究和行业关注于让这类系统自我修复，但当前的技术要求软件组件简单、无状态和仅限崩溃。使用遗留的企业信息系统很难达到这些要求，此外，当前的方法只考虑一小部分非常通用的自动恢复操作。为了解决这些局限性，需要准确的故障定位和诊断方法，以及复杂的错误恢复算法。这项研究的目标是设计这种新的算法和方法，并将它们集成到系统管理基础设施中，以创建一个能够提供端到端错误检测、故障定位和诊断以及错误恢复的自愈元服务器，并在数据中心规模上这样做。我们的方法是为系统的正常运行创建可伸缩的运行时行为模型，并识别系统行为和运行时模型预测之间的偏差。任何此类异常通常都是由工作负载更改、系统更改或系统故障引起的；我们可以监视工作负载和系统更改，因此它们都不会被识别、假定故障并启动恢复。考虑到企业信息系统运行成本巨大，我们的研究有望在降低成本方面具有非常重要的意义。

Enterprise information systems today comprise large, complex interconnected assemblages of software, computers, networks, etc. These systems are mission critical, and as such are expect to be highly available, offer good performance, and operate within constrained budgets. Currently, this is achieved by having highly available datacenter administrators actively monitor critical services, identifying failures, service-level agreement (SLA) violations, etc., localizing and diagnosing faulty components, and generally ensuring that the system is in a correctly functioning state. This approach is expensive (the cost to business in the US is approximately $100 billion annually, rising 30% by 2018; US Bureau of Labor Statistics). While there has been recent research and industry attention given to having such systems self-heal, current techniques require the software components to be simple, stateless, and crash-only. These requirements are hard to attain with legacy enterprise information systems, Furthermore, current approaches consider only a small set of very generic automated recovery actions. To address these limitations, accurate fault localization and diagnosis methods, together with sophisticated error-recovery algorithms, are required. It is the objective of this research to devise such novel algorithms and methods, and to integrate them within a systems-management infrastructure, so as to create a self-healing meta-server which will provide end-to-end error detection, fault localization and diagnosis, and error recovery, and to do so at the datacenter scale. Our approach is to create scalable, run-time behavioural models of the normal operation of the system, and to identify deviations between system behaviour and run-time-model prediction. Any such anomalies are typically caused by workload change, system change, or system fault; we can monitor workload and system change, and so neither of those are identified, a fault is presumed, and recovery initiated. Given the large costs of running enterprise information systems, our research is expected to be very significant in cost reduction.