Labeling Schemes for Improving Cilksan Performance

Abstract

While race detection algorithms like SP-bags have provably good theoretical properties, large overheads exist in practice, which urges the need for performance optimization. In this thesis, I propose labeling schemes as a method of circumventing many of the expensive operations in Cilksan, an implementation of the SP-bags algorithm. The proposed labeling schemes give strands of a parallel program labels during the execution of Cilksan, allowing Cilksan to shortcut the processing of certain memory accesses if the label comparison allows. I describe and prove correctness for two labeling schemes, the procedure labeling scheme and the prefix labeling scheme, implement both in Cilksan, and measure their performance. While the results show that the overhead of maintaining labels is too high in my implementation, the labeling schemes manage to circumvent many of the memory access operations, suggesting the merit of a more performant implementation of the same schemes.