ITR: Automatic On-The-Fly Detection, Characterization, Recovery and Correction of Software Bugs in Production Runs

生产运行中软件错误的自动即时检测、表征、恢复和纠正摘要我们建议开发一个全面的调试系统，用于生产运行的自动即时调试。我们的系统处理调试的所有方面，包括错误检测、表征、恢复和更正。该系统紧密集成了计算机硬件、操作系统、数据挖掘和编译器支持方面的创新。更具体地说，我们建议的系统依赖于以下创新：(1)低开销的编译器指导的检查，以检测错误，并将信息传递给数据挖掘器、操作系统和硬件，以表征错误；(2)低开销的数据挖掘算法，其构建和使用模型来检测、表征和纠正错误；(3)新的硬件支持，以非常低的开销并且对用户透明地回滚和确定地重新执行有错误的代码段；以及(4)操作系统支持回滚和重新执行硬件不支持的代码段。这四个层面紧密结合在一个软件原型中。我们的工作直接旨在改善IT革命历史上的黑点：程序员工作效率低下。我们可以通过专注于调试生产运行来应对软件调试挑战，这要归功于在硬件中提供了非常低的开销错误表征支持。一旦检测到错误，额外的编译器、数据挖掘器和操作系统层可以增强硬件的能力。我们希望这项工作将产生广泛的影响，因为加快调试过程可以大幅提高IT专业人员和学生的生产率。此外，开发的思想还可以用于其他异常检测，如入侵检测和安全攻击，这对我们的社会非常重要。

Automatic On-the-fly Detection, Characterization, Recovery and Correction of Software Bugs in Production RunsAbstractWe propose to develop a comprehensive debugging system for automatic on-the-fly debugging of production runs. Our system addresses all the aspects of debugging, including bug detection, characterization, recovery and correction. The system tightly integrates innovations in computer hardware, operating system, data mining, and compiler support. More specifically, our proposed system hinges on the following innovations: (1) low-overhead compiler-directed checking to detect bugs and pass information to the data miner, operating system, and hardware to characterize the bugs; (2) Low-overhead data mining algorithms that build and use models for bug detection, characterization, and correction; (3) Novel hardware support to roll back and deterministically re-execute buggy sections of code with very low overhead and transparently to the user; and (4) operating system support to roll back and re-execute code sections that cannot be supported by hardware. These four layers are tightly integrated in a software prototype.Our work directly aims at improving what has historically been the dark spot of the IT revolution: poor programmer productivity. We can meet the software debugging challenge by focusing on debugging production runs, thanks to providing very low-overhead bug characterization support in hardware. The additional compiler, data miner, and operating system layers can enhance the power of the hardware once the bug is detected. We hope the work will have a broad impact, since expediting the debugging process can lead to dramatic increases in the productivity of IT professionals and students. In addition, the ideas developed can also be used for other anomaly detection, such as intrusion detection and security attacks, which is very important to our society.