Scalable Monitoring Technique for Detecting Races in Parallel Programs

Detecting races is important for debugging shared-memory parallel programs, because the races result in unintended nondeterministic executions of the programs. Previous on-they techniques to detect races have a bottleneck caused by the need to check or serialize all accesses to each shared variable in a program that may have nested parallelism with barrier synchronization. The new scalable monitoring technique in this paper reduces the bottleneck signi cantly by checking or serializing at most 2(B + 1) non-nested accesses in an iteration for each shared variable, where B is the number of barrier operations in the iteration. This technique, therefore, makes on-they race detection more scalable.