Interleaving generation for data race and deadlock reproduction
Luis M. Carril, Walter F. Tichy · 2015
Concurrency errors, like data races and deadlocks, are difficult to find due to the large number of possible interleavings in a parallel program. Dynamic tools analyze a single observed execution of a program, and even with multiple executions they can not reveal possible errors in other reorderings. This work takes a single program observation and produces a set of alternative orderings of the synchronization primitives that lead to a concurrency error. The new reorderings are enforced under a happens-before detector to discard reorderings that are infeasible or do not produce any error report. We evaluate our approach against multiple repetitions of a state of the art happens-before detector. The results show that through interleaving inference more errors are found and the counterexamples enable easier reproducibility by the developer.