Energy Consumption and Scalability Evaluation for Software Transactional Memory on a Real Computing Environment

Transactional Memory is a concurrent programming abstraction that overcomes several of the limitations found in traditional synchronization mechanisms. As it is a more recent abstraction, little is known about energy consumption of Software Transactional Memories (STM). In this context, this work presents an analysis and characterization of energy consumption and performance of four Transactional Memory libraries: TL2, TinySTM, SwissTM, and AdaptSTM, using the STAMP benchmarks. Although most works in the state-of-the-art chose to evaluate Transactional Memories through simulation, in this work the benchmarks are run in actual computers, avoiding the known issues with modeling power consumption in simulators. Our results show that SwissTM is the most efficient library of the four in terms of energy consumption and performance for the default configurations, followed by AdaptSTM, TinySTM, and TL2, for most of the execution scenarios and 8 threads at most. STM's scalability is directly tied to the strategies for detection and resolution of conflicts. In this perspective, AdaptSTM is the best STM for applications with short transactions, SwissTM presents the best results for medium transactions, and long transactions with medium/high contention are best handled by TL2. On the other hand, TinySTM shows the worst scalability for most scenarios, but with good results for applications with very small abort rates.