CAREER: Analyzing Distributed Systems Behavior Using Repeated Execution

现代分布式系统极其复杂，在很大程度上是由于单个节点的复杂性、节点的不可靠性和异步性以及不可预测的网络消息延迟和排序。使这些系统的开发更加复杂的是，这些系统存在多个可能不兼容的版本，并且需要建立正确的系统，也表现出高性能。以前的测试和模拟框架的特点要么是大量的人工工作，要么是自动搜索二元决策问题的违规-存在或不存在错误。我们正在开发自动化和交互技术，以帮助开发人员了解分布式系统实现的行为。通过利用演进的框架，并以结构化、直截了当的方式检测实现，我们正在构建专注于理解系统行为的开发工具，而不仅仅是识别正确的错误。这一重点的改变将使更通用的工具能够在测试生产率之外提高开发效率。这项研究从三个方面进行：1)使用重复执行的数据挖掘来开发自动化工具，以提取执行行为和性能；2)开发适用于用例描述和自动化处理的灵活的执行描述，允许用户与他们的工具进行更直观的交互；3)将测试工具与修订控制系统相结合，实现多版本分析和长期进度跟踪。完成后，这项研究将减少开发人员设计、更新和调试分布式系统所需的工作量，并可能激发新一代系统调试器的重新创建，这些调试器不仅分析正确性，而且分析性能和复杂性。

Modern distributed systems are extremely complex, due in large part toindividual node complexity, node unreliability and asynchrony, andunpredictable network message delays and orderings. Further complicatingdevelopment of these systems is both the presence of multiple potentiallyincompatible versions of the systems, and the need to build correct systemsalso exhibiting high performance. Prior testing and simulation frameworks arecharacterized either by extensive manual effort, or automated search forviolations of a binary decision problem---the presence or absence of a bug. We are developing automated and interactive techniques for helping developersunderstand the behavior of distributed systems implementations. By leveragingthe evolved frameworks, and instrumenting implementations in structured,straightforward ways, we are building development tools focused onunderstanding system behavior rather than merely identifying correctnesserrors. This change in focus will enable more general tools that improvedevelopment productivity in addition to testing productivity.This research is proceeding on three fronts: 1) developing automated tools using data mining with repeated executions to extract execution behaviors andperformance, 2) developing flexible execution descriptions suitable for bothuse-case descriptions and automated processing, allowing more intuitiveinteraction between users and their tools, and 3) incorporating testing toolswith revision control systems, enabling multi-version analysis and long-termprogress tracking. When completed, this research will reduce developer effortnecessary to design, update, and debug distributed systems, and may inspirecreation of a new class of systems debuggers analysing not just correctness,but also performance and complexity.