Lightweight Transactional Arrays for Read-Dominated Workloads

Abstract

Many common workloads rely on arrays as a basic data structure on top of which they build more complex behavior. Others use them because they are a natural representation for their problem domains. Software Transactional Memory (STM) has been proposed as a new concurrency control mechanism that simplifies concurrent programming. Yet, most STM implementations have no special representation for arrays. This results, on many STMs, in inefficient internal representations, where much overhead is added while tracking each array element individually, and on other STMs in false-sharing conflicts, because writes to different elements on the same array result in a conflict. In this work we propose new designs for array implementations that are integrated with the STM, allowing for improved performance and reduced memory usage for read-dominated workloads, and present the results of our implementation of the new designs on top of the JVSTM, a Java library STM.