RaceHunter Dynamic Data Race Detector

Abstract

Data races represent a class of concurrency errors when two threads access a shared memory location without proper synchronization. Data races are hard to reveal and debug. This paper presents RaceHunter, a dynamic data race detection technique, which monitors executions of shared-memory concurrent programs, discovers pairs of conflicting memory accesses, and systematically checks them for data races. RaceHunter does not report false data races when the target software exploits nonstandard synchronization primitives or unknown synchronization protocols, and it can also detect data races missed by other approaches. Dynamic data race detectors can monitor continuous, e.g., real-life, program executions or they can verify relatively short program executions, e.g., organized by system tests. The latter is the primary use case scenario for RaceHunter.