Review of Nesting in Transactional Memory

In face of the multicore paradigm that is now standard in most machines, we are left with the challenge of how to take advantage of all the potential parallelism: Programming applications that explore concurrency is far from trivial. It is in this context that the Transactional Memory (TM) abstraction promises to simplify this task. So far, each part of an application identified as atomic is typically a sequential set of instructions. However, these may account for a large part of the program which is thus running sequentially and may represent bottlenecks in the system. These atomic blocks could be executed faster if they contained latent parallelism that could be used efficiently. In the context of Transactional Memories, however, there has been little work to allow a transaction to be split in several parts, all executed concurrently. Such behavior may be attained through parallel nested transactions. Besides providing a more flexible TM, using parallel nested transactions may actually yield better performance in applications resorting to TM for synchronizing access to shared data. In this work we describe how we may still reduce the time to execute a highly-contending write dominated workload despite the inherent conflicts that arise in it. We expect to provide an adaptation of a lock-free and multi-version STM, the JVSTM, to use a parallel nesting model that allows the aforementioned flexibility.