Deterministic OpenMP for Race-Free Parallelism

Amittai Aviram and Bryan Ford
Yale University

3rd USENIX Workshop on Hot Topics in Parallelism (HotPar '11),
May 26-27, 2011, Berkeley, California, USA


Recent deterministic parallel programming models show promise for their ability to replay computations and reproduce bugs, but they currently require the programmer to adopt restrictive or unfamiliar parallel constructs. Deterministic OpenMP (DOMP) is a new deterministic parallel environment built on the familiar OpenMP framework. By leveraging OpenMP's block-structured synchronization annotations, which are largely compatible with the constraints of a deterministic model, DOMP eases the parallelization of legacy serial code and preserves substantial compatibility with OpenMP software. A few OpenMP constructs, however, such as atomic and critical, are semantically nondeterministic and unsupported in DOMP. In three well-known parallel benchmark suites, we find that most (81%) uses of these nondeterministic constructs express programming idioms that are compatible with determinism but not directly expressible in OpenMP. DOMP therefore adds new OpenMP constructs to express such idioms deterministically, supporting pipelines and generalized reductions.

Paper: PDF

This research is sponsored by the National Science Foundation under grant CNS-1017206.