Using Transactional Memory on CC-NUMA systems

Transactional Memory (TM) research has focused on multi-core processors and SMP machines; limited research has been aimed at the clusters, leaving the area of CC-NUMA (Cache Coherent Non-Uniform Memory Access) systems unexplored. The CC-NUMA system's memory is physically distributed which brings the different latency between local and remote memory access. The TMs designed for shared memory architecture are not suitable for CC-NUMA's distributed memory architecture. To be aware of the different memory access latency, we introduce a layered conflict detection strategy corresponding to the system's layer. The layered strategy provides a good practical performance on CC-NUMA systems.