A hardware/software framework for supporting transactional memory in a MPSoC environment

Manufacturers are focusing on multiprocessor-system-on-a-chip (MPSoC) architectures in order to provide increased concurrency, rather than increased clock speed, for both large-scale as well as embedded systems. Traditionally lock-based synchronization is provided to support concurrency; however, managing locks can be very difficult and error prone. In addition, the performance and power cost of lock-based synchronization can be high. Transactional memories have been extensively investigated as an alternative to lock-based synchronization in general-purpose systems. It has been shown that transactional memory has advantages over locks in terms of ease of programming, performance and energy consumption. However, their applicability to embedded multi-core platforms has not been explored yet. In this paper, we demonstrate a complete hardware transactional memory solution for an embedded multi-core architecture, consisting of a cache-coherent ARM-based cluster, similar to ARM's MPCore. Using cycle accurate power and performance models for the transactional memory hardware, we evaluate our architectural framework over a set of different system and application settings, and show that transactional memory is a promising solution, even for resource-constrained embedded multiprocessors.