Race Condition Detection for Debugging Shared-memory Parallel Programs

This thesis addresses theoretical and practical aspects of the dynamic detecting and debugging of race conditions in shared-memory parallel programs. To reason about race conditions, we present a formal model that characterizes actual, observed, and potential behaviors of the program. The actual behavior precisely represents the program execution, the observed behavior represents partial information that can be reasonably recorded, and the potential behavior represents alternate executions possibly allowed by nondeterministic timing variations. These behaviors are used to characterize different types of race conditions, general races and data races, which pertain to different classes of parallel programs and require different detection techniques. General races apply to programs intended to be deterministic; data races apply to nondeterministic programs containing critical sections. We prove that, for executions using synchronization powerful enough to implement two-process mutual exclusion, locating every general race or data race is an NP-hard problem. However, for data races, we show that detecting only a subset of all races is sufficient for debugging. We also prove that, for weaker types of synchronization, races can be efficiently located. We present post-mortem algorithms for detecting race conditions as accurately as possible, given the constraint of limited run-time information. We characterize those races that are direct manifestations of bugs and not artifacts caused by other races, imprecise run-time traces (causing false races to appear real), or unintentional synchronization (caused by shared-memory references). Our techniques analyze the observed behavior to conservatively locate races that either did occur or had the potential of occurring, and that were unaffected by any other race in the execution. Finally, we describe a prototype data race detector that we used to analyze a sample collection of parallel programs. Experiments indicate that our techniques effectively pinpoint non-artifact races, directing the programmer to parts of the execution containing direct manifestations of bugs. In all programs analyzed, our techniques reduced hundreds to thousands of races down to four or fewer that required investigation.