On the design space of Parallel Nesting

One of the key strengths of Transactional Memory (TM) is that transactions compose. Yet, very few TM implementations allow a transaction to be split into several parts that execute concurrently. To overcome this limitation in transaction’s composability, TMs must support parallel nesting. In this work we elaborate on how the design choices of the underlying TM are inherently coupled with the complexity bounds achievable in the support for parallel nesting. For that, we compare three different parallel nesting algorithms that explore opposite directions in the design space. We also show that the theoretical analysis of these algorithms may be deceiving and we support that conclusion with empirical results in known benchmarks.