Thinking Beyond Race Conditions

Abstract

Multi-threaded parallel programs perform accesses to shared variables, which require application of a synchronization strategy. Absence of synchronization among threads may lead to error situations, such as data races. Data races, which involve concurrent accesses to a single shared variable, are categorized as low-level. Synchronization strategies to alleviate lowlevel data races do not guarantee freedom from data races on a higher-level of abstraction. In this paper, we discuss some scenarios where accesses to a group of variables may result in data races. Further, we discuss a method to statically detect such situations and describe its integration into Bauhaus tool suite.