Efficient Dynamic Analysis for Detecting and Tolerating Program Anomalies

商业、交通、经济和科学都越来越依赖于软件。为了满足这些需求，程序员已经转向托管语言，如Java、Ruby和c#，因为由于垃圾收集和类型安全，它们有助于更快地生成更高质量的软件。TOIBE发现，大多数新软件都在使用Java或其他托管语言。尽管这些语言减少了错误，但程序仍然包含内存和语义错误。NIST估计，软件错误每年给美国经济造成590亿美元的损失。由于大多数先前用于查找bug的工具会因因素和/或增加大量内存开销而减慢程序速度，因此它们通常只在测试期间或崩溃后使用，而不是在生产期间使用。本研究将探索用于检测、报告和容忍错误的新方法，这些方法在已部署的软件中足够有效。这些技术包括新的、低开销的编译器分析、运行时检测和运行时算法。一个关键的特性是非常低的内存和适度的时间开销，这使得它们在部署软件中非常有吸引力，特别是在进程争夺内存资源的多核环境中。一个独特的方面是，垃圾收集器将汇总堆中的数据结构，以发现程序模式和异常。建议的工作将检测、报告和容忍内存泄漏，否则将使程序崩溃。这个项目的研究影响将是更高质量的软件，bug更少，即使有一些错误也能保持正确执行，通过保持软件运行帮助用户的工具，以及帮助开发人员发现、诊断和修复错误的工具。研究者将使这些工具公开可用，培训研究生，并指导增加少数民族对计算机科学的参与。

Businesses, transportation, economies, and science all increasingly depend on software. To manage these demands, programmers have turned to managed languages such as Java, Ruby, and C#, because they help produce higher quality software faster due to garbage collection and type safety. TOIBE fnds that most new software is using Java or other managed languages. Although these languages reduce errors, programs still contain memory and semantic errors. NIST estimates that software errors cost the US economy $59 billion a year. Because most prior tools for finding bugs slow programs down by factors and/or add substantial memory overheads, they are typically employed only during testing or after a crash, not during production.This research will explore new approaches for detecting, reporting, and sometimes tolerating errors that are efficient enough to use in deployed software. The techniques include new, low overhead compiler analyses, runtime instrumentation, and runtime algorithms. A key feature will be very low memory and modest time overheads that make them appealing for use in deployed software, especially in multicore where processes are competing for memory resources. A unique aspect will be that the garbage collector will summarize data structures in the heap to discover program patterns and anomalies. The proposed work will detect, report, and tolerate memory leaks that would otherwise crash the program. The research impact from this project will be higher quality software with fewer bugs that keeps executing correctly in spite of some errors, tools that help users by keeping software running, and tools that help developers find, diagnose, and fix errors. The investigator will make the tools publicly available, train graduate students, and mentor to increase minority participation in computer science.