Improving Systems Reliability Through Record/Replay

软件错误会严重影响软件可靠性，导致服务中断、数据丢失和安全漏洞。不幸的是，大多数生产软件故障都没有得到诊断。这种情况的发生有两个原因。首先，广泛使用的软件崩溃的绝对数量例如，微软在最近的一份出版物中报告了每天超过100万次崩溃，Mozilla的工作人员向我们报告说，他们每天看到数万次崩溃。其次，目前的崩溃报告技术，如核心转储和堆栈跟踪，提供的上下文非常少。因此，自动分组(即由同一个错误导致的崩溃集)、诊断(即这是并发错误、缓冲区溢出等)和确定错误的优先顺序(即某个错误是否可能被利用)一直是不切实际的。为了解决这个问题，我们一直在构建一个工具链，允许开发人员记录和重放生产应用程序，称为Castor。卡斯特有几个目标。首先，我们正在尝试以足够低的性能和空间开销为应用程序记录/回放提供一个平台，使其在生产中默认开启成为现实，并在错误发生时捕获它们。在之前的工作中，我们已经证明，即使是要求最苛刻的多核科学和服务器工作负载，也可以记录下来，而CPU开销只有几个百分点。展望未来，我们希望展示以适度的空间开销实现这一目标的新技术，并探索这样做可以实现的错误分类和诊断的新方法。此研究提案的主要目标是改进开发人员工具，使所有崩溃都能得到诊断。我们希望在我们的记录/重放工作的基础上继续发展，在可用性、性能和空间效率方面提高其核心能力，并将其与动态分析、程序切片和其他技术结合使用，以开发更好的工具来自动分组、诊断错误并确定错误的优先级。次要目标是探索记录/重放的替代用途。首先，我们将研究为任何性能分析工具生成足够精确的周期重放的技术。其次，我们希望研究如何帮助开发人员自动创建支持容忍非故障停止错误的容错流程。与此相关，我们可以扩展思想，使容错软件在几乎没有开销的情况下对漏洞具有健壮性。这个研究计划将有助于我更广泛的研究目标，即在一个越来越依赖计算机的世界中提高系统可靠性。我们将有机会对软件的开发和维护方式产生影响。它还将为HQP在加拿大需求旺盛的领域提供丰富的培训场地。

Software bugs significantly hamper software reliability leading to service outages, data loss, and security vulnerabilities. Unfortunately, most production software failures go undiagnosed. This occurs for two reasons. First, the sheer number of crashes in widely used software for example, Microsoft reported seeing over a million crashes a day in one recent publication and staff at Mozilla reported to us they see tens of thousands of crashes a day. Second, current crash reporting technologies, such as core dumps and stack traces, provide very little context. Consequently, automatically grouping (i.e., the set of crashes caused by the same bug), diagnosing (i.e., this is a concurrency bug, buffer overflow, etc.), and prioritizing bugs (i.e., whether a bug is potentially exploitable) has been impractical.To address this, we have been building a toolchain that allows developers to record and replay production applications, called Castor. Castor has several goals. First, we are attempting to provide a platform for application record/replay with low enough performance and space overhead to make it practical to leave on by default in production, capturing bugs when they occur. In prior work, we have shown that it is possible to record even the most demanding multi-core scientific and server workloads with CPU overheads of a few percent. Moving forward, we hope to demonstrate new techniques to achieve this with modest space overheads and explore new approaches to bug triage and diagnosis that this can enable.The primary objective of this research proposal is to improve developer tools such that no crash is left undiagnosed. We want to continue building on our record/replay work, improving its core capabilities in terms of usability, performance and space efficiency, and to use it in conjunction with dynamic analysis, program slicing, and other techniques to develop better tools for grouping, diagnosing and prioritizing bugs automatically.A secondary objective is to explore alternative uses of record/replay. First, we will study techniques to generate a cycle-accurate replay good enough for any performance analysis tools. Second, we want to investigate helping developers automatically create fault tolerant processes that support tolerating non-fail-stop errors. Related to this we can extend the ideas to make fault tolerant software robust to exploits with little to no overhead.This research program will contribute to my broader research goal of improving systems reliability in a world ever more dependent on computers. We will have the opportunity to make an impact in the way software is developed and maintained. It will also provide a rich training ground for HQP in a field that is in high demand in Canada.