Enabling Speculative Parallelization via Merge Semantics in STMs

STM (Software Transactional Memory) systems can be used to speculatively parallelize irregular applications such as those based on graphs and trees. While the transactional paradigm is suitable for such speculative parallelization, STMs do not deal with the semantics of speculation. In this paper we introduce merge semantics for speculations. STMs optimistically execute code and monitor the execution for memory access conflicts. On detecting a conflict between a pair of transactions the STM performs a rollback on one of them, discarding all the work that has been done until that point. This wastage leads to performance and scalability issues when there are many transactional conflicts. The key insight of this paper is that it is sometimes possible to salvage partially completed work and merge it with the other transaction. In this paper we motivate the need for the merge construct and develop the ideas behind it. We propose a simple API which can be used to define the merge operation. We discuss applications which would benefit from this construct. We implement the supporting framework and demonstrate its benefits on a Connected Components benchmark and a Minimum Spanning Tree benchmark. We report performance improvements of more than 75% when compared to a traditional STM parallelization of the Minimum Spanning Tree benchmark. Our framework also demonstrates excellent scalability when there are a large number of conflicts, a scenario where traditional STM systems do not scale well.