Efficient Debugging of Parallel Programs

本项目涉及并行程序调试系统的设计和实现的技术、算法。反排分析的使用提供了程序执行中事件之间因果关系的信息，而无需重新执行程序进行调试。反排分析直接使程序员关注问题的原因。解决了两个主要问题：(1)通过开发一种显着减少程序执行期间产生的开销的方法使反排分析变得实用；(2)将反排分析的语义扩展到并行程序。增量跟踪使用对已调试程序的语义分析，使得在执行过程中只生成少量跟踪就可以进行跟踪分析。反排分析扩展到应用于并行程序，处理同步活动，进程之间的依赖关系和共享内存访问。本文还描述了一种在合作过程的相互作用中检测竞争条件的方法。有效地调试并行和顺序程序所必需的技术和算法被整合到一个建立在商用多处理器上的原型系统中。

This project addresses techniques algorithms for design and implementation of a debugging system for parallel programs. The use of flowback analysis provides information on causal relationships between events in a program's execution, without having to re-execute the program for debugging. Flowback analysis directly focuses the programmer on the cause of a problem. Two major issues are addressed: (1) making flowback analysis practical by developing a method that significantly reduces the overhead induced during program execution, and (2) extending the semantics of flowback analysis to parallel programs. Incremental tracing, using semantic analyses of the debugged program, makes the folwback analysis practical with only a small amount of trace generated during execution. Flowback analysis is extended to apply to parallel programs, handling synchronization activities, dependencies between processes, and shared memory access. A method to detect race conditions in the interactions of the co-operating processes is also described. The techniques and algorithms necessary to efficiently debugging both parallel and sequential programs are incorporated in a prototype system built on a commercially available multiprocessor.