Practical Parallel Nesting for Software Transactional Memory

Transactional Memory (TM) provides a strong abstraction to tackle the challenge of synchronizing concurrent tasks that access shared state. Yet, most TMs do not allow a single transaction to contain parallel code. We propose an efficient parallel nesting algorithm to explore existing latent parallelism within a transaction. If this intra-transaction parallelism has reduced conflict probability (compared to the inter-transaction parallelism), then it may be worthy to execute less transactions at a given time, but have each one parallelized and using several available cores. We provide practical support for parallel nesting in the first lock-free parallel nesting algorithm with support for multi-versions. Our prototype builds over an available multi-version TM, which we outperform on standard benchmarks by up to 2.8x. We show improvements over parallel nesting alternatives of up to 3.6x.