Runtime Behavioural Models for Dependable Systems

现代计算包括数据中心、客户端设备和连接它们的互联网。由于数以百万计的服务器、数十亿台客户端设备和众多网络都处于不同的管理控制之下，实现可靠的计算是困难的。即使硬件和软件都是高度可靠的，有如此多的应用程序，在任何给定的时间，它们中的许多都会出现故障。数据中心的可靠性是通过实施面向恢复计算(ROC)的复制和运行时软件管理来实现的。网络可靠性是通过网络接口和互连的冗余来实现的。实现可靠的客户端更加困难，因为它们没有冗余。它们运行多个应用程序，因此数据中心ROC中使用的通用恢复操作不起作用。目前，客户端记录设备和应用程序日志数据，并将其发送到数据中心进行分析。这样做的问题是，日志记录往往是开发人员认为可能有助于调试应用程序的任何内容，而应用程序通常对用户或操作员没有用处；很难根据日志记录确定适当的恢复操作；更糟糕的是，开发人员可能错过了关键的日志点；最后，发送用于分析的日志数据量巨大且不断增长，公司每天收集TB级的日志记录。我们计划扩大我们在系统监控和错误检测方面的工作。目前，我们混合使用回归分析、信息论模型和基于日志消息文本相似性的聚类来对软件系统的运行时行为进行建模。我们将监视的行为与我们的模型进行比较；与模型的偏离表示工作负载更改、软件更改或系统中的错误。当我们可以消除前两者时，就可以推断错误。在这项研究中，我们将扩展我们的运行时模型的范围，以纳入拟议的自我恢复操作对系统行为的影响。当前的工具通过应用有限的自动化选项序列来解决错误。大多数软件系统有许多针对不同情况的恢复选择，这在由各种软件供应商的支持组织维护的知识数据库中是显而易见的。当这样的替代方案可用时，使用我们的建模方法的故障定位和诊断可能是至关重要的，因为固定的恢复选项序列将是无效和昂贵的。可能的恢复操作集将从支持知识数据库中已知的成功技术中提取。在应用任何恢复操作后，我们希望密切监控系统行为，以确定操作的成功或失败，并根据需要进行调整。除了为客户创建运行时管理系统外，我们还希望开发自动日志点插入技术，以解决从操作员和用户的角度来看日志记录质量较差的问题。

Modern computing comprises data centers, client devices, and the Internet connecting them. With millions of servers, billions of client devices, and numerous networks, all under different administrative control, achieving dependable computing is difficult. Even if both hardware and software are highly reliable, with so many applications, at any given time numerous of them will have failed. Data-center dependability is achieved by replication and untime software management implementing Recovery-oriented Computing (RoC). Network dependability is achieved via redundancy of network interfaces and interconnections. Achieving dependable clients is more difficult, as they do not have redundancy. They run multiple applications, and so generic recovery actions used in data-center RoC do not work. Currently clients record device and application log data and send it to a data center for analysis. The problem with this is that log records tend to be whatever a developer thought might help in debugging an application which is often not useful for a user or an operator; determining appropriate recovery actions based on log records is hard; worse, the developer may have missed critical log points; finally, the amount of log data sent for analysis is vast and growing, with companies gathering terabytes of log records daily. We plan to extend our work in system monitoring and error detection. Currently we model the run-time behaviour of software systems using a mixture of regression analysis, information-theoretic models, and clustering based on textual similarity of logs messages. We compare monitored behaviour with our models; deviation from the model indicates workload change, software change, or error in the system. When we can eliminate the first two, the can infer error. In this research we will extend our range of run-time models to incorporate the effects of proposed self-recovery actions on system behaviour. Current tools address errors by applying a limited sequence of automated options. Most software systems have numerous recovery alternatives for different situations, as is evident in the knowledge databases maintained by support organizations of various software vendors. When such alternatives are available, fault localization and diagnosis using our modeling approach is likely critical, as a fixed sequence of recovery options will be ineffective and costly. The set of possible recovery actions will be extracted from known-successful techniques from support knowledge databases. After applying any recovery action, we expect to closely monitor system behaviour to determine the success or failure of the action, and adjust as necessary. In addition to creating a runtime management system for clients, we expect to develop automated logpoint insertion techniques to address the poor quality of log records as viewed from the operator and user perspective.